---
kind: concept
slug: data-augmentation
title: Data augmentation
summary: >-
  Extra training signal from labelled data: transform examples while keeping labels valid, mix patches with label weights, generate images, or label recombined pairs with a teacher. Choose by label type, scarcity and what you can change.
tags: [training, data, computer-vision, graph-learning, few-shot-learning, nlp, generative-models, time-series, class-imbalance, tabular-data, speech, lidar, semi-supervised]
papers: ["2012.07177", "1802.09567", "2103.00111", "1811.04241", "2009.08369", "2101.06395", "2201.09206", "2202.06709", "2110.04994", "1706.06480", "1804.02386", "2006.01791", "2010.08240", "2304.08466", "2010.11683", "2011.11156", "2104.10858", "2107.12283", "2605.11989", "2401.06853", "2408.15857", "2609.10524", "1708.09427", "1807.11809", "2606.23286", "2106.14112", "2606.29836", "2111.01243", "2512.06714", "2607.19153", "1505.03046", "1810.05726", "2106.01342", "2110.09348", "2205.06743", "2212.08104", "2304.02858", "2605.30099", "2609.37870", "2002.04688", "2010.14945", "2110.06207", "2212.07784", "1702.00871", "2106.05819", "2512.13100", "2411.11394", "2412.17226", "1704.02581", "2012.09311", "2408.06083", "1802.00752", "2104.07713", "2110.01852", "2201.03299", "2204.08610", "2405.05258", "2609.24410"]
related: [dataset-curation, few-shot-learning, contrastive-learning, time-series-forecasting, generative-adversarial-network, language-model-pretraining, linear-probing, vision-transformer, semantic-segmentation, semi-supervised-learning, self-training, domain-adaptation, diffusion-models, classifier-free-guidance, adversarial-robustness, named-entity-recognition, sequence-labeling, knowledge-distillation, ensembles, transfer-learning, self-supervised-learning, time-series-classification, physically-based-rendering, graph-neural-network, salient-object-detection, imitation-learning, robot-simulation, vision-language-action-model, multimodal-language-model, hallucination-mitigation, vision-language-pretraining, object-detection, monocular-depth-estimation, image-manipulation-detection, test-time-augmentation, dropout]
status: draft
updated_at: 2026-11-01
---

## Overview

Labelled data is expensive, and a model often sees too few examples of some
classes or situations. Data augmentation adds training signal without new
labelling. Most variants transform existing examples. The invariant they all
need is that the transformed example still has a correct label, either the
original one or one that has been updated to match the transform
[arxiv:2012.07177 §3] [arxiv:1802.09567 §4.2]. One variant works on
embeddings rather than pixels, and there the invariant is that sampled vectors
stay near the class's estimated distribution [arxiv:2101.06395 §3.2.2].

Several variants depart from the transform pattern. Augmented SBERT builds new
sentence pairs and takes their labels from a teacher model, the generative
variant makes new images from class-name prompts, and the model-filtered
sentence variant keeps candidate sentences that a tagger agrees with
[arxiv:2606.29836 §3.3]. Two more generate or insert data: a per-type GAN
makes synthetic failure rows for tabular sensor data, and a forecaster inserts
a virtual value between each pair of real readings [arxiv:2607.19153 §3.2.5]
[arxiv:2512.06714 §3.5]. A survey also describes language models writing
training examples [arxiv:2111.01243 §5]. Two further papers generate whole
training items: a multimodal model writes navigation instructions for routes
built from house-tour videos, and a diffusion model makes LiDAR objects that
are then placed in scenes [arxiv:2411.11394 §3.2] [arxiv:2412.17226 §3.2].
They are included because their
papers present them as augmentation or training-data generation, and their
sections say where the invariant differs.

The contrastive variants (graph, time-series and tabular pre-training) use no
label in their pre-training step. Their invariant is that two views of one
sample should agree [arxiv:2106.14112 §3.3] [arxiv:2103.00111 §6]
[arxiv:2106.01342 §4].

The variants on this page differ in what they change:

- whole images composited from parts of two images, with object masks rewritten
  to match [arxiv:2012.07177 §3];
- a patch cut from one image and pasted onto another at a saliency peak, with
  the label weights set by the patch area [arxiv:2006.01791 §3.1] [arxiv:2006.01791 §3.2];
- individual images or characters, with flips, colour inversion or a fixed set
  of geometric and photometric transforms [arxiv:1802.09567 §4.2]
  [arxiv:1811.04241 §4] [arxiv:2009.08369 §4.1]. A survey of 40 agricultural
  studies counts rotations in 12 papers and gives no parameter ranges
  [arxiv:1807.11809 §4.5];
- patches cut from large microscope images, rotated by 90, 180 and 270 degrees,
  with the sampling stride set per class [arxiv:1706.06480 §4.3.4];
- feature vectors sampled from a class Gaussian that is borrowed from similar
  base classes [arxiv:2101.06395 §3.2.2];
- repeated copies of scarce satellite images, each with its own random
  transform [arxiv:2201.09206 §3.3];
- graph structure and node features, used to make the views that contrastive
  pre-training compares [arxiv:2103.00111 §6.1];
- a weakly and a strongly perturbed copy of each raw time series, used as the
  two views of a contrastive pretraining step [arxiv:2106.14112 §3.1];
- a bundle of strong image-level policies used when training vision
  transformers from scratch [arxiv:2202.06709 app. A];
- sentence pairs recombined from the labelled set and scored by a cross-encoder
  [arxiv:2010.08240 §3.1];
- images drawn from a fine-tuned text-to-image model, one prompt per class
  [arxiv:2304.08466 §4.1];
- tokens in a tagged sentence, replaced or shuffled so that the BIO tags stay
  valid [arxiv:2010.11683 §3];
- sentences from unlabelled abstracts, tagged by a model and kept only when
  the model finds every known entity in them [arxiv:2606.29836 §3.3];
- patch tokens of a vision transformer, each with a dense label from an offline
  annotator, mixed as patch embeddings [arxiv:2104.10858 §3.3];
- pixels of a segmentation model, mixed in the loss only, with unblended label
  maps [arxiv:2107.12283 §6];
- offline rotated and noisy copies of a small photo set [arxiv:2605.11989 §3.3.3];
- perturbed copies of synthetic temporal graphs, with facts dropped, relations
  renamed, entities renamed and years shifted [arxiv:2401.06853 §3.2.2];
- mosaic and mixup as named steps of detector training, with no parameters
  given [arxiv:2408.15857 §3.1.1] [arxiv:2408.15857 §3.2];
- label-safe transforms drawn on the fly for whole mammograms: flips, rotations,
  shear, zoom and channel shifts [arxiv:1708.09427 §3.1.2];
- reflections, shifts, scales and shears that expand a small phone-photo set
  nine-fold before the split [arxiv:2606.23286 §5.1];
- flips and rotations that grow a grain photo set sixfold [arxiv:2609.10524 §4.1];
- synthetic failure rows from one WGAN-GP per failure type, pooled with the real
  minority rows [arxiv:2607.19153 §3.2.5];
- one virtual value inserted between each pair of consecutive demand readings
  [arxiv:2512.06714 §3.5];
- training examples that a language model writes, as a survey describes them
  [arxiv:2111.01243 §5];
- a heavy chain of rotation, zoom, shear, colour and brightness changes, and
  flips, for an ImageNet-pretrained classifier on field photos
  [arxiv:1810.05726 § Deep Learning];
- CutMix on table cells and mixup on the row embeddings, as views for
  contrastive pre-training [arxiv:2106.01342 §4];
- resampling of minority rows, by SMOTE and its variants, ADASYN, random
  over- or under-sampling, or a GAN, before a classifier is fitted
  [arxiv:2304.02858 §4.1] [arxiv:2304.02858 tab. 1].
- virtual raindrops ray-traced onto real clean street photos, with the drop
  mask taken from the simulation [arxiv:2609.37870 §3].
- a batch-level transform pipeline in a deep learning library, run on the GPU,
  with rotation, zoom and lighting defaults [arxiv:2002.04688 §2.1];
- node and edge drops for graph contrastive learning, with probabilities set
  by node centrality so that influential structure survives
  [arxiv:2010.14945 §3.3.1] [arxiv:2010.14945 §3.3.2];
- RandAugment and label smoothing inside a long closed-set training recipe,
  used for open-set recognition [arxiv:2110.06207 app. D];
- cached Mosaic and MixUp in detector training, followed by a final stage with
  large-scale jittering [arxiv:2212.07784 §3.3] [arxiv:2212.07784 tab. 7];
- a second frame made by warping an annotated still image along a smooth flow
  field, so that video-style training pairs exist without filming
  [arxiv:1702.00871 §4];
- edge drops whose probabilities come from a learned network, trained to make
  two views of a graph differ, with a penalty on how many edges go
  [arxiv:2106.05819 §3.2];
- robot demonstrations re-rendered with another arm and gripper in the same
  pose, with the background filled in behind the original arm
  [arxiv:2512.13100 §4.1] [arxiv:2512.13100 §4.2];
- route-and-instruction pairs, one per house-tour route, with the instruction
  written by a multimodal model and checked against the route's room and action
  labels [arxiv:2411.11394 §3.2];
- LiDAR foreground objects from a text-and-box diffusion model, placed in a
  range-image scene model with a per-category loss [arxiv:2412.17226 §3.2]
  [arxiv:2412.17226 §3.3];
- random 3D rotation, scaling and shear of skeleton joint coordinates, applied
  to each training sequence [arxiv:1704.02581 §4.3];
- a random gamma-type brightness curve on synthetic indoor images, with the
  depth map left unchanged [arxiv:2408.06083 §3.1];
- real face crops stitched onto other identities inside elastically deformed
  masks, which are known because the method makes them
  [arxiv:2012.09311 §3.2].

Creating new labelled scenes from 3D meshes is a related but different
method. It is covered under scope below. Test-time augmentation averages
predictions over transformed inputs and changes no training data. It is not
a training method, so it is covered under scope below too.

## Variants

### Copy-Paste compositing for instance segmentation

Two random training images are each scale-jittered and flipped. A random
subset of the objects from one image is pasted onto the other, and the
objects of the receiving image that are covered are trimmed or removed from
the labels [arxiv:2012.07177 §3]. Paste locations are random and no scene
context is modelled [arxiv:2012.07177 §3]. Card: [Ghiasi et al. (Copy-Paste)](/papers/2012.07177).

Reported gains on COCO for Mask R-CNN-style models are +1.0 to +1.5 box AP and
+0.5 to +1.1 mask AP across seven models [arxiv:2012.07177 tab. 1]. With
self-training on top, the strongest model reaches 57.3 box AP and 49.1 mask AP
on test-dev [arxiv:2012.07177 tab. 4]. The gain is largest when labels are
scarce: on 10% of COCO it adds 6.9 box AP over scale jittering alone and 4.8
over large scale jittering [arxiv:2012.07177 §4.3].

The paper's ablation finds that pasting a random subset beats pasting a single
object or all objects [arxiv:2012.07177 tab. 10]. Blending the paste edges
made no difference to AP [arxiv:2012.07177 tab. 10]. Random pairing often
produces scenes that do not match, such as indoor and outdoor images mixed in
about 47% of generated examples [arxiv:2012.07177 app. C] [arxiv:2012.07177 tab. 12].

### Saliency-guided patch mixing for classification

A patch is cut from a source image and pasted at the same location on a target
image. The patch is centred on the most salient pixel of the source, found with
an OpenCV saliency detector, and the two labels are mixed in proportion to the
patch area [arxiv:2006.01791 §3.1] [arxiv:2006.01791 §3.2] [arxiv:2006.01791 §3.3].
Among five placement schemes, pasting at the corresponding location gives the
lowest error in the paper's experiments [arxiv:2006.01791 §3.4]. The motivation
is that a random patch often lands on background, and the mixed label would
then credit background to the source class [arxiv:2006.01791 §1].
Card: [Uddin et al. (SaliencyMix)](/papers/2006.01791).

For WideResNet-28-10 without traditional augmentation, top-1 error falls from
5.18% to 4.04% on CIFAR-10 and from 23.21% to 19.45% on CIFAR-100, against
CutMix [arxiv:2006.01791 tab. 1]. With crop and flip added, the gap is much
smaller: 2.87% against 2.76% on CIFAR-10 and 16.66% against 16.56% on CIFAR-100
[arxiv:2006.01791 tab. 1]. On ImageNet, ResNet-50 moves from 21.40% to 21.26%
top-1 over CutMix [arxiv:2006.01791 tab. 2]. The ImageNet numbers are the best
run per method, while the CIFAR numbers are means over five runs, so they are
not directly comparable [arxiv:2006.01791 §4.1.2] [arxiv:2006.01791 tab. 1].

The released code differs from the paper in ways that change the augmentation.
It computes one saliency box per batch, from the first partner image, and
pastes that box from every partner [code: SaliencyMix_CIFAR/saliencymix.py#L241-L242].
Its box side is the square root of 1 − λ times the image side, and it recomputes
the label weight from the clipped area [code: SaliencyMix_CIFAR/saliencymix.py#L195-L197]
[code: SaliencyMix_CIFAR/saliencymix.py#L243]. The ImageNet script trains on a
200-class Tiny-ImageNet path [code: SaliencyMix-ImageNet/train.py#L79-L80],
so the ImageNet-1k numbers need more than the released script.

### Rotated patches with class-balanced sampling for pixel-wise segmentation

For pixel-wise segmentation of large microscope images, the training inputs
are 1000 × 1000 px patches cut from the full image with a sliding window
[arxiv:1706.06480 §4.2] [arxiv:1706.06480 §4.3.3]. Each patch is rotated by 90,
180 and 270 degrees [arxiv:1706.06480 §4.3.4]. The window stride is set per
class so that each class yields about the same number of patches, which gives
rarer classes a smaller stride [arxiv:1706.06480 §4.3.4]. The exact stride
values are not stated. The worked example uses a 100 px stride on a
7000 × 8000 px image, which gives 4331 patches [arxiv:1706.06480 §4.3.4].
Card: [Azimi et al. (steel microstructure)](/papers/1706.06480).

The paper's ablation adds the two steps one at a time. On SEM images with a
fine-tuned fully convolutional network, balancing alone moves pixel accuracy
from 92.01% to 92.11%, and rotation on top of balancing moves it to 93.92%
[arxiv:1706.06480 tab. 5]. Per-class pixel accuracy does not move in one
direction: balancing alone lowers tempered martensite from 71.43% to 57.44%,
and rotation brings it back to 72.62% [arxiv:1706.06480 tab. 6]. At the object
level, the full method reaches 93.94% from 90.97% with balancing alone
[arxiv:1706.06480 tab. 2]. The text gives the augmentation gain as 2 points
[arxiv:1706.06480 §6]; the table difference is 2.97 points.

The rotation is a label-preserving assumption. The paper reports similar
responses to rotated patches and calls the system rotation invariant
[arxiv:1706.06480 §6], but it does not test invariance of the architecture. All
numbers come from one image-level split of 21 images, with no repeated runs
[arxiv:1706.06480 §5.1]. Use this only where a 90-degree turn leaves the
label unchanged.

### Feature-space sampling for few-shot classifiers

This variant does not change images at all. The feature extractor is frozen.
For each support feature, the method borrows a Gaussian from the two nearest
base classes, averages it with the support feature's own statistics, and
draws 750 vectors per class from it [arxiv:2101.06395 §3.2.2] [arxiv:2101.06395 §3.2.3].
A logistic regression is trained on the support features plus the drawn
vectors [arxiv:2101.06395 §3.2.3]. Card: [Yang et al. (distribution calibration)](/papers/2101.06395).

Before sampling, the features are passed through a power transform with
exponent 0.5 [arxiv:2101.06395 §3.2.2]. The method assumes non-negative
features, such as pooled post-ReLU features [arxiv:2101.06395 §4.1.3].

In the authors' miniImageNet 5-way ablation, the two steps add to each other
for 1-shot accuracy [arxiv:2101.06395 tab. 4]:

| Power transform | Generated features | 1-shot (%) | 5-shot (%) |
|---|---|---|---|
| no | no | 56.37 | 79.03 |
| yes | no | 64.30 | 81.33 |
| no | yes | 63.70 | 82.26 |
| yes | yes | 68.57 | 82.88 |

The CUB value of α in the paper is not the one the released code uses, and
the paper's tieredImageNet results are not covered by that code
[arxiv:2101.06395 §4.1.3] [arxiv:2101.06395 §4.1.1]. Treat the cross-dataset
numbers as the authors' results only.

### Repeated augmented copies of scarce classes

The University-1652 benchmark has one satellite image per building, so
training sees a few satellite images for each class. The fix is to draw each
satellite image k times in a training step. Every copy gets its own random
transform, such as shifting, padding, cutting or colour change
[arxiv:2201.09206 §3.3]. Card: [Dai et al. (FSRA)](/papers/2201.09206).

With the triplet loss alone, drone-to-satellite AP rises as the copy count
goes up [arxiv:2201.09206 tab. 7]:

| Copies k | D→S AP (%) | S→D AP (%) |
|---|---|---|
| 1 | 84.85 | 81.52 |
| 2 | 86.36 | 82.69 |
| 3 | 86.71 | 83.37 |

The extra copies make training slower but leave inference unchanged
[arxiv:2201.09206 §3.3] [arxiv:2201.09206 §4.4]. The best k depends on the
batch size: the paper reports k = 3 at batch 8 and k about 5 at batch 16, and
the batch-16 values appear only in a figure [arxiv:2201.09206 app. B].

### Graph augmentations for contrastive self-supervision

For graphs, a view is a perturbed copy of a node, subgraph or whole graph.
Attribute-level views mask or shuffle node features. Topology-level views
modify edges, apply graph diffusion, or sample a subgraph [arxiv:2103.00111 §6.1].
The model is trained so that two views of the same object agree and views of
different objects do not [arxiv:2103.00111 §6]. Card: [Liu et al. (graph SSL survey)](/papers/2103.00111).

The survey reports that contrastive methods using augmented views (MVGRL,
MERIT, SubG-Con) outperform generation-based and auxiliary-property methods
on node classification [arxiv:2103.00111 §8] [arxiv:2103.00111 tab. 5]. It
also notes that augmentation choices in contrast-based methods are tuned
empirically and are costly to tune [arxiv:2103.00111 §3.2]. The augmentation
strengths used by the surveyed methods are not stated [arxiv:2103.00111 §8].

### Centrality-adaptive drops for graph contrastive learning

Two views of a graph are built by removing edges and masking feature
dimensions. Each edge and each feature dimension gets a removal probability
from a score: edges from the centrality of their endpoints, feature
dimensions from the centrality-weighted sum of the nodes that have them
[arxiv:2010.14945 §3.3.1] [arxiv:2010.14945 §3.3.2]. Low-scoring items are
removed more often, and a cap keeps any item from being almost always removed
[arxiv:2010.14945 eq. 4]. Degree, eigenvector and PageRank centrality are the
three measures tried [arxiv:2010.14945 §3.3.1]. Card:
[Zhu et al. (GCA)](/papers/2010.14945).

The uniform version of the same two steps is the baseline. On five node
classification datasets the adaptive scheme beats uniform dropping in every
case, by 0.1 to 1.6 accuracy points, and the largest gap is on
Amazon-Computers [arxiv:2010.14945 tab. 4]. The paper's own reading is that
each adaptive part helps and that both together help most [arxiv:2010.14945 §4.3].
The rules and the cap are described in the paper, but the official code
differs from the paper in edge weighting and in the PageRank update, so
reimplementations should check [arxiv:2010.14945 §3.3.1]. Those differences are
recorded on the card.

### Learned edge drops for graph contrastive learning

Here the drop probabilities are not set by a rule. A graph neural network
(the augmenter) outputs one drop logit per edge. The encoder is trained to
match each graph with its edge-weighted copy, and the augmenter is trained to
make those copies differ. A penalty on the mean drop probability per graph
keeps the augmenter from dropping nearly everything [arxiv:2106.05819 §3.2]
[arxiv:2106.05819 eq. 8]. Card: [Suresh et al. (AD-GCL)](/papers/2106.05819).

The paper argues that the min-max form links the objective to a graph
information bottleneck, so the encoder keeps less information that the task
does not need [arxiv:2106.05819 §3.1]. The theorem behind this assumes a
1-WL-powerful encoder and a countable graph space, and the paper gives no
guarantee for a trained network [arxiv:2106.05819 §3.1] [arxiv:2106.05819 app. B].
Gradients through the discrete edge choice use a Gumbel-max relaxation, so
the drop is soft during training [arxiv:2106.05819 §3.2].

The authors' unsupervised results are linear-probe accuracies over 10 seeds.
On the TU collection the paper counts significant gains over InfoGraph and
GraphCL on 5 of 8 and on 6 of 8 datasets [arxiv:2106.05819 §5.1]. On RDT-M5K
the value is 54.93 for AD-GCL-OPT against 53.05 for GraphCL
[arxiv:2106.05819 tab. 1]. The fixed-λ and tuned-λ versions of the method
give similar numbers, which the paper reads as robustness to the penalty
weight [arxiv:2106.05819 §5.1]. On transfer to chemistry and biology tasks,
AD-GCL-FIX has a mean rank of 2.4 across nine datasets, with the baseline
numbers copied from earlier papers [arxiv:2106.05819 §5.3]
[arxiv:2106.05819 app. G].

The regulariser sets the drop ratio. For classification, λ from 0.1 to 10
gives expected drop ratios from about 80% down to under 10%, and accuracy is
flat across that range [arxiv:2106.05819 §5.2] [arxiv:2106.05819 fig. 3]. For
regression, the paper finds a smaller penalty, and so more dropping, works
better [arxiv:2106.05819 app. F]. Edge dropping is a proof-of-concept search
space; the paper says other families may do better [arxiv:2106.05819 §3.2].

The released code differs from the paper in several details, among them how
dropped edges are handled, the contrastive temperature and the default dropout
[arxiv:2106.05819 app. G]. Those differences change the augmentation, so check
the card's Paper vs code section before reproducing the numbers.

### Weak and strong views for time-series contrastive pretraining

Time series have no labels in the pretraining stage, so the augmentation only
decides which two views the model must match. TS-TCC makes a weak view with
jitter and scaling. It makes a strong view by cutting the series into a random
number of segments, shuffling them, and adding jitter. The paper says the
segment count should grow with sequence length, and that jitter should be
smaller for normalised data [arxiv:2106.14112 §3.1]. The two views feed a
temporal contrasting module and a contextual contrasting loss
[arxiv:2106.14112 §3]. Card: [Eldele et al. (TS-TCC)](/papers/2106.14112).

The ablation on the three labelled tasks shows the two views doing different
work. On HAR, the weak view alone gives 76.55% accuracy and the strong view
alone gives 60.23%, against 90.37% for both [arxiv:2106.14112 tab. 4]. On
Sleep-EDF the single views give 80.90% and 78.55%, against 83.00%. On
Epilepsy the single views give 97.18% and 97.14%, against 97.23%. The text
reads this as one view alone hurting HAR and Sleep-EDF, but not Epilepsy
[arxiv:2106.14112 §5.4]. All figures are author-reported over five seeds
[arxiv:2106.14112 §4.2].

The linear probe on the frozen encoder reaches 90.37% on HAR, against 90.14%
for the supervised model. On Sleep-EDF it reaches 83.00% against 83.41%
supervised, and on Epilepsy 97.23% against 96.66% [arxiv:2106.14112 tab. 2].
Strong jitter strength is not stated in the paper, and the released code uses
different segment counts and scales from the text (see the card's Paper vs
code). The released code also samples the augmented views once per run, so
each sample keeps the same two views in every epoch [code: dataloader/dataloader.py#L32-L33].
Re-sampling per epoch would be a different method. Read from the code, the
strong permutation may give every channel the same ordering on multichannel
data [code: dataloader/augmentations.py#L43]. Check this before reusing the
strong view on data such as HAR's nine channels.

### Strong views matched through similarity distributions

Keep the weak view as a teacher, and ask the strong view to reproduce the
weak view's similarity profile over a queue of stored keys, rather than to
match its own key [arxiv:2104.07713 §3.2] [eq. 6]. The term is added to a
standard contrastive loss with weight β = 1.0, with τ = 0.2 and key momentum
0.999 [arxiv:2104.07713 eq. 7] [arxiv:2104.07713 §3.3]. The strong view is
built from 14 operations, drawn in five rounds with probability 0.5 each
[arxiv:2104.07713 §3.3] [tab. 1]. Card: [Wang & Qi (CLSA)](/papers/2104.07713).
Author-reported ImageNet linear top-1 at 200 epochs is 69.4% against 67.5% for
MoCo v2 [arxiv:2104.07713 tab. 2]. Without the distribution term, a strong
query alone gives 67.7 and a strong query with a strong key gives 67.0
[arxiv:2104.07713 tab. 5]. The wall-clock cost is 70 h against 53 h for MoCo v2
at 200 epochs [arxiv:2104.07713 tab. 6].

### Strong augmentation in vision-transformer training

Training a vision transformer or a hybrid from scratch on small data often
uses a bundle of strong image-level policies. The paper names RandAugment,
Random Erasing, label smoothing, mixup, CutMix and stochastic depth, with
AdamW and a cosine schedule [arxiv:2202.06709 app. A]. Card: [Park & Kim (ViT analysis)](/papers/2202.06709).
Most values for these components are not stated in the paper, so a
reimplementation has to choose them [arxiv:2202.06709 app. A].

The paper's own analysis gives mixed effects. Strong augmentation reduces the
size of the Hessian eigenvalues but produces many negative ones
[arxiv:2202.06709 app. F]. Models trained without it are over-confident on
CIFAR-100, and models trained with it become under-confident
[arxiv:2202.06709 app. F]. The card records no accuracy ablation of this
bundle on its own [arxiv:2202.06709 app. A].

### RandAugment and label smoothing in a long closed-set recipe

For open-set recognition, the paper trains a closed-set classifier with a
strong recipe and then scores unknowns with the maximum logit. The recipe
includes RandAugment, with its count and magnitude tuned per dataset, and
label smoothing on TinyImageNet only, at 0.9 [arxiv:2110.06207 app. D]. In the
official code the smoothing value is the mass taken away from the true class,
so the true class keeps only 0.1 of the target [arxiv:2110.06207 app. D]. On
TinyImageNet, switching from crop and flip to RandAugment at 200 epochs adds
1.8 AUROC points, and adding label smoothing at 600 epochs adds 4.6 more
[arxiv:2110.06207 app. C]. These are single-setting ablations on five
known/unknown splits. Card:
[Vaze et al. (open-set recognition)](/papers/2110.06207).

### Recombined sentence pairs with teacher labels

Sentence-pair models can learn from pairs that nobody labelled by hand. A slow,
accurate cross-encoder is trained on the gold pairs. It then scores new pairs
made by recombining gold sentences, and a fast bi-encoder is trained on the gold
pairs plus these soft-labelled pairs [arxiv:2010.08240 §3.1]. The cross-encoder's
score is the label, so the bi-encoder learns from soft targets [arxiv:2010.08240 §3.1]
[arxiv:2010.08240 §7]. Card: [Thakur et al. (Augmented SBERT)](/papers/2010.08240).

Which pairs are sampled decides the gain. The paper compares random sampling,
kernel density estimation (KDE) that matches the gold score distribution,
BM25 lexical neighbours, semantic-search neighbours, and a merge of the last
two [arxiv:2010.08240 §3.1]. On the five in-domain datasets, the best strategy
beats the lower-bound SBERT baseline by about 1.0 to 6.0 points. Random sampling
falls below that baseline on Spanish-STS, at 62.05 against 72.07
[arxiv:2010.08240 tab. 5]. BM25 gives the smallest silver set in every row, for
example 3,964 pairs on Spanish-STS against 230,364 for KDE [arxiv:2010.08240 tab. 11].

For domain adaptation, no labelled target data is needed. The cross-encoder is
trained on a labelled source domain and labels the unlabelled target pairs
[arxiv:2010.08240 §3.2]. The largest reported change is Quora to Sprint, where
AUC(0.05) rises from 0.505 to 0.875 [arxiv:2010.08240 tab. 6]. In most other
rows the result stays below the in-domain upper bound [arxiv:2010.08240 tab. 6].

This is closer to pseudo-labelling than to transforming an example. The new
pairs are not transformed copies of gold pairs, and their validity rests on the
teacher's scores being useful labels. It is listed here because the paper
presents itself as data augmentation.

### Generated images from a fine-tuned text-to-image model

Instead of transforming existing images, this variant makes new labelled images
with a generator. The authors fine-tune Imagen on the ImageNet-1K training
images, prompt it with each class name, and generate as many images per class
as there are real ones, at scale factors from 1 to 10
[arxiv:2304.08466 §4.1] [arxiv:2304.08466 §4.3]. The label is the class in the
prompt, so label validity depends on the generator producing that class.
Card: [Azizi et al. (synthetic ImageNet)](/papers/2304.08466).

Classifiers trained only on generated images do worse than real-only training
in every row of the classifier table. A ResNet-50 reaches 64.96% top-1 on real
validation images with generated data at 256 × 256, and 69.24% at 1024 × 1024,
against 73.09% for real data under the CAS recipe [arxiv:2304.08466 tab. 2].
Adding 1.2M generated images to the real training set raises ResNet-50 top-1
from 76.39% to 78.17% [arxiv:2304.08466 tab. 3].

The gain does not grow with the synthetic set. In the scaling table, the best
total size is 2.4M images at both resolutions, with +1.22 at 256 and +1.73 at
1024. At 6.0M total the 256 result is 0.30 below the real-only baseline, and
the 1024 result is 0.05 below [arxiv:2304.08466 tab. 4]. The table's size axis
is read as the total training set, since its 1.2M row equals the real-only
baseline; the paper does not define the column [arxiv:2304.08466 tab. 4].

Fine-tuning used 256 TPU-v4 chips for 210K base steps and 490K steps for the
64-to-256 super-resolution stage, at batch 2048 [arxiv:2304.08466 §4.1]. The
generation cost of the 1.2M to 12M sets is not reported [arxiv:2304.08466 §4.3].
The paper links no code, and its generator is not linked from the paper either
[arxiv:2304.08466 §4.1].

### Per-type GAN oversampling for tabular failure data

This variant synthesises minority rows for tabular sensor data. Instead of one
generator for all failures, it trains one per failure type. Failure rows are
split by type (HDF, PWF, OSF, TWF), a WGAN-GP is fitted to each group with its
own Optuna study, and each generator draws samples in proportion to its group's
size. A classifier then keeps only the synthetic rows it labels as real with
high confidence. The paper uses LightGBM for this filter by default, and the
survivors are pooled with the real rows [arxiv:2607.19153 §3.2.5]
[arxiv:2607.19153 §3.3.1]. Card:
[Lazanas & Kampouropoulos (multi-generator GAN)](/papers/2607.19153).

On the AI4I 2020 dataset, with a Random Forest, the pooled per-type generators
reach PR-AUC 0.8574 and recall 0.8824 at precision 0.5696. The cost-sensitive
forest reaches 0.835 and 0.5882, a single GAN over all failures 0.8245, and
SMOTE 0.7525 [arxiv:2607.19153 tab. 4]. The paper reports one run per
configuration, so the 0.02 PR-AUC gap over the cost-sensitive baseline is not
separated from run-to-run variation [arxiv:2607.19153 §4.1]. The GAN objective
is the forest's validation PR-AUC, and the paper does not say whether the SVM
scenarios reuse the tuned generators [arxiv:2607.19153 §3.3.2]. The single-GAN
run took about 11 h 20 min and the multi-GAN run about 14 h 5 min, against
minutes for the classical scenarios [arxiv:2607.19153 §5.1].

### Virtual values between samples for demand forecasting

Water-demand forecasts have large errors where demand changes sharply between
two 15-minute readings [arxiv:2512.06714 §1]. The EDCGRU variant inserts one
virtual value between each pair of consecutive readings and classifies it with
the same k-means step as a real reading. The input window doubles from 96 to
192 rows [arxiv:2512.06714 §3.5] [arxiv:2512.06714 §4.2]. The authors report
one virtual value per gap as the best of the values they tried, and say that
more did not help [arxiv:2512.06714 §4.2]. Card:
[Salloom et al. (water demand)](/papers/2512.06714).

The paper never gives the rule that computes a virtual value. It says only that
the values relieve nonlinearity [arxiv:2512.06714 §3.5], so an implementation
has to choose one, such as linear interpolation, and report the choice. On two
district metering areas from one data set, one-step MAPE falls from 1.26% to
0.98% and from 1.31% to 0.99% against the same model without virtual values,
and training time rises from 192 s to 850 s on the first area
[arxiv:2512.06714 tab. 6]. The abstract's 30% error reduction does not name the
metric it refers to [arxiv:2512.06714 §7]. The k-means class features that this
model also uses are an input feature, not augmentation, and are not covered
here.

### Language-model-written training examples (survey only)

A survey of large pre-trained language models describes two ways a model can
write data for another model. The first adds labelled or silver-labelled
examples to the training set, with filtering of noise. The second writes
auxiliary text such as counterexamples, clarifying questions, contexts or
inference rules that help another model explain or support its output
[arxiv:2111.01243 §5.1] [arxiv:2111.01243 §5.2]. Card:
[Min et al. (pre-trained LM survey)](/papers/2111.01243).

The survey runs no experiments and gives no result for either use, so this page
can describe the method but not its effect [arxiv:2111.01243 §5]. For measured
results with model-labelled data, see the recombined-pair and model-filtered
variants above.

### Token-level replacement for sequence labelling

Tagging tasks such as named-entity recognition give every token a label, so
changing a word can change its label. This variant adapts four simple methods
to BIO tags, and none of them needs a translation or parsing model
[arxiv:2010.11683 §3]. Label-wise token replacement swaps a token for another
token that has the same full BIO tag. Synonym replacement swaps a token for a
WordNet synonym, and the extra words take I- tags when the token began a
mention. Mention replacement swaps a whole mention for another mention of the
same type and rewrites its tags to fit. Shuffle within segments permutes the
tokens inside each mention or each run of outside tokens, and leaves the tags
in place [arxiv:2010.11683 §3]. Label-wise replacement and shuffling keep the
label sequence unchanged, while synonym and mention replacement can change it
[arxiv:2010.11683 tab. 1]. Card: [Dai & Adel (NER augmentation)](/papers/2010.11683).

On the 50-sentence training subsets, every single method is marked significant
against no augmentation, for both taggers and both data sets
[arxiv:2010.11683 tab. 3]. For the BiLSTM tagger on MaSciP, the 50-sentence
score rises from 53.0 to between 58.8 and 60.8 F1 [arxiv:2010.11683 tab. 3].
Applying all four methods gives the largest average gain over the eight
columns: +3.3 F1 for the LSTM and +1.8 for SciBERT [arxiv:2010.11683 tab. 3].
The gain is concentrated in the small sets. On the full training set several
single-method cells are equal to or below the baseline; SciBERT on i2b2-2010
ranges from 87.2 to 87.8 against 87.8 without augmentation
[arxiv:2010.11683 tab. 3] [arxiv:2010.11683 §4.4]. The released code runs only
the BERT-based tagger, so the LSTM results cannot be reproduced from it (see
the card's Paper vs code).

### Model-filtered sentences from unlabelled abstracts

The authors of a scientific-entity tagger say that easy text augmentations
risk changing entity spans, so they use a model-based procedure instead
[arxiv:2606.29836 §3.3]. Candidate sentences come from unannotated abstracts
and must contain a known entity. A tagger trained on the training set labels
each candidate. A candidate is kept only when the tagger finds every known
entity it matches, and the kept sentences are added until the augmented set is
as large as the training set [arxiv:2606.29836 §3.3]. Card:
[Zhang et al. (NLP technology entities)](/papers/2606.29836).

On the authors' own 311-sentence test split, the SciBERT plus BiLSTM tagger
goes from 86.49 to 87.00 F1 with this augmentation [arxiv:2606.29836 tab. 7].
This is one model run, with no standard deviation reported [arxiv:2606.29836 §4.1].
The gain of 0.51 F1 is therefore not separated from run-to-run variation. The
repository has no script for the procedure, so it has to be rebuilt from §3.3
[arxiv:2606.29836 §3.3].

This is closer to pseudo-labelling than to transforming an example, like the
recombined-pair variant above. The new labels come from the model's own
output, filtered by agreement with gold entities. Check the labels on a held-out
split before you trust the added data.

### Token labeling with MixToken for vision transformers

The class token of a vision transformer is the only token that receives the
classification loss, so the patch tokens get no direct supervision
[arxiv:2104.10858 §1]. Token labeling gives each patch token its own target.
The targets come from a dense label map that an offline annotator computes once
for every training image, here an ImageNet-only NFNet-F6 classifier
[arxiv:2104.10858 §3.2] [arxiv:2104.10858 §4.1]. Training adds a token-level
cross-entropy term to the class term [arxiv:2104.10858 §3.2]. MixToken is a
CutMix-style mix on patch embeddings: a box of tokens is exchanged between two
images, and the class target is mixed by the share of tokens from each image
[arxiv:2104.10858 §3.3] [arxiv:2104.10858 eq. 7]. Card:
[Jiang et al. (token labeling)](/papers/2104.10858).

The annotator is the costly step. Its top five logits per cell were stored in
half precision, which the authors put at about 10 GB, against about 2 TB for
full maps [arxiv:2104.10858 §4.1]. The paper says the added compute is
negligible but gives no wall-clock time [arxiv:2104.10858 §3.2].

Author-reported results: removing MixToken drops LV-ViT-S from 83.3% to 81.3%
top-1 [arxiv:2104.10858 tab. 3]. With the same ReLabel labels, MixToken scores
83.0 against 82.8 for CutMix [arxiv:2104.10858 tab. 2]. A weaker annotator,
EfficientNet-B3 at 81.6% top-1, still helps LV-ViT-S, while NFNet-F6 gives 83.3
[arxiv:2104.10858 §4.2] [fig. 4]. The released code weights the token loss at
1.0, where the paper uses 0.5, and it does not mix token targets under
MixToken [arxiv:2104.10858 §3.2] [arxiv:2104.10858 eq. 6]. Check the card's
Paper vs code before copying either.

### Mixup on the loss only, for segmentation

A building segmenter mixes two training images with a small weight λ, but it
does not blend their label maps. It runs the network once on the blended image,
then takes the cross-entropy against each label map, weighted by λ and 1 − λ
[arxiv:2107.12283 §6] [arxiv:2107.12283 eq. 5] [arxiv:2107.12283 eq. 6]. Mixup
is applied to the cross-entropy term only, not to the Tversky term
[arxiv:2107.12283 §6]. The same pipeline uses random 448 px crops from 600 px
tiles, flips, rotations, and random changes to brightness, hue, saturation and
contrast [arxiv:2107.12283 §6]. Card: [Sirko et al. (building detection)](/papers/2107.12283).

The paper uses λ = 0.05 [arxiv:2107.12283 §6]. With that value the partner's
label map gets 95% of the loss weight [arxiv:2107.12283 eq. 6]. The paper does
not say how partner images are chosen or how λ is sampled [arxiv:2107.12283 §6].
The authors kept the labels separate after preliminary runs found that blended
labels did worse once pixel weights were used [arxiv:2107.12283 §6]. The
abstract reports +0.12 mAP for mixup, while the ablation appears only as a
figure [arxiv:2107.12283 fig. 8]. The same pipeline's self-training gives random
black masks on the student's inputs, so that it learns to handle no-data
regions [arxiv:2107.12283 §8].

### Many colour and crop views pooled over frozen features

No backbone is trained. Each slide is recoloured 50 times, and each recolouring
gives 20 random square crops that a frozen ImageNet encoder embeds
[arxiv:1802.00752 §2.3]. The haematoxylin and eosin magnitudes are scaled by
two independent draws from U(0.7, 1.3) [arxiv:1802.00752 §2.3]. The crop
features of one recolouring are pooled with a power mean (p = 3)
[arxiv:1802.00752 eq. 1], and a LightGBM classifier learns from the pooled
descriptors [arxiv:1802.00752 §2.5]. Card: [Rakhlin et al. (breast histology)](/papers/1802.00752).
Author-reported accuracy is 87.2% for four classes under 10-fold
cross-validation on 400 images [arxiv:1802.00752 tab. 1]. The code also
rotates crops by multiples of 90° and flips them, which the paper does not
mention [code: feature_extractor.py#L163-L171].

### Offline rotated and noisy copies of a small photo set

A transfer-learning benchmark expanded 654 smartphone photos into 8,502 images
before training. Each photo is rotated by 45, 135, 225 and 315 degrees, and each
rotated copy gets noise drawn uniformly up to 10, 25 or 50 percent of the pixel
range [arxiv:2605.11989 §3.3.3]. The paper does not say whether the noise is
added to or replaces pixel values, how the corners left by rotation are filled,
or whether the augmentation runs before or after the train/validation split
[arxiv:2605.11989 §3.3.3]. If it runs first, copies of one photo can land on
both sides of the split, which leaks information into validation. Card:
[Abou Baker et al. (transfer learning)](/papers/2605.11989).

The results are averaged over the five datasets [arxiv:2605.11989 §5.1], so the
effect of this augmentation alone is not reported.

### Noisy-student loop with audio augmentation

A teacher pseudo-labels one unlabelled batch at a time, and a larger student is
trained on the pseudo-labelled data, the clean labelled data and augmented
copies [arxiv:2609.24410 §8.2] [fig. 8]. The augmentations are Gaussian noise,
time stretch and pitch shift, applied at every generation
[arxiv:2609.24410 §6.2]. Their magnitudes and selection are not stated, and no
pseudo-label filter is described [arxiv:2609.24410 §6.2] [arxiv:2609.24410 §7.3].
Card: [Safieh et al. (Jordanian speech)](/papers/2609.24410). The authors'
evaluation WER falls from 0.568 to 0.515 over five generations
[arxiv:2609.24410 tab. 4]. Each generation also uses a larger model, and the
card's results contain no run that isolates the augmentation.

### Perturbed temporal graphs for reasoning fine-tuning

The reasoning model of TG-LLM is trained on verified temporal graphs, but at
test time it reads graphs that its own first stage predicted. The authors
treat that gap as a risk and add four perturbations to each training example:
drop facts that appear in neither the question nor the chain of thought,
replace relation names with synonyms, rename entities with one global map, and
shift every year by one offset. The same change is applied to the question,
chain of thought, candidates and answer, so the example stays consistent
[arxiv:2401.06853 §3.2.2] [arxiv:2401.06853 app. B]. Card: [Xiong et al. (TG-LLM)](/papers/2401.06853).

In the authors' TGQA ablation for Llama2-13B, the augmentation raises exact
match from 0.723 to 0.782 without external knowledge, and from 0.753 to 0.797
with it [arxiv:2401.06853 tab. 5]. These are single runs. The paper's human
error table shows that not every error type falls: the temporal-graph error
rate rises from 0.03 to 0.04 when augmentation is added
[arxiv:2401.06853 tab. 3].

The code appends one augmented copy of each training example, so the set
doubles, and it draws the entity map and the year offset once per run
[arxiv:2401.06853 §3.2.2] [code: src/SFT_TG_Reasoning.py#L97-L124]. It sets no
random seed, so two runs get different augmented data unless you fix one
yourself [code: src/SFT_TG_Reasoning.py#L101-L105]. The source data is built
from YAGO11k and GPT-3.5 outputs; check its terms before reusing the generated
set [arxiv:2401.06853 §2].

### Mosaic and mixup in detector training

The YOLOv8 review names two training augmentations. Mosaic combines "four or
more" images into one training sample, and mixup blends images
[arxiv:2408.15857 §3.1.1] [arxiv:2408.15857 §3.2]. The paper gives no canvas
size, placement rule, box update, mixup ratio or per-step probability, so
neither step can be reimplemented from the paper alone [arxiv:2408.15857 §3.2].
It runs no experiment that isolates either one, and its claim that the
augmentation helps small objects comes without figures [arxiv:2408.15857 §8].
Card: [Yaseen (YOLOv8 review)](/papers/2408.15857). Measure these steps on your
own data before relying on them.

### Cached Mosaic and MixUp, then large-scale jittering

For a real-time one-stage detector, the paper keeps a cache of decoded images
and draws Mosaic and MixUp samples from it. Stage one of training uses these
two steps. The last 20 epochs switch to large-scale jittering, which the paper
presents as a fine-tuning stage close to the real image distribution
[arxiv:2212.07784 §3.3] [arxiv:2212.07784 tab. 1]. A large cache with random
eviction behaves like uncached augmentation. A small cache with first-in,
first-out eviction makes the same image reappear with different augmentations
in one batch [arxiv:2212.07784 §3.3]. Cached Mosaic takes 24.0 ms per 100
images against 87.1 ms uncached [arxiv:2212.07784 tab. 7]. Card:
[Lyu et al. (RTMDet)](/papers/2212.07784).

On COCO, the two-stage schedule with cached Mosaic and MixUp reaches 51.3 AP
for RTMDet-l, against 46.7 with large-scale jittering alone and 49.8 with
uncached Mosaic and MixUp in both stages [arxiv:2212.07784 tab. 7]. The
ablations are single runs with no spread reported. The paper's numbers are
for its own detector family, so they do not transfer directly to other models.

### Flips and rotations, split before or after

A grain-variety classifier grows about 4,500 captured images to 27,000 with
horizontal and vertical flips and rotations of 45, 90 and 180 degrees
[arxiv:2609.10524 §4.1] [arxiv:2609.10524 fig. 2]. The paper does not say
whether this happens before or after the train/test split. Table 2 gives 22,500
training and 4,500 test images, and the test count equals the per-class count
before augmentation. That suggests a split on originals, but the paper does not
say so [arxiv:2609.10524 tab. 2]. If the split came after augmentation, copies of
one grain could sit on both sides, so test accuracy would overstate the result
[arxiv:2609.10524 §4.1]. The reported base-model accuracies of 97.5 to 99.2 percent
are validation figures, and the results tables give no run without augmentation
[arxiv:2609.10524 tab. 4]. Card: [Islam et al. (rice variety)](/papers/2609.10524).

### Heavy transform chain for a fine-tuned classifier

Two ImageNet-pretrained networks, ResNet-50 and Inception-v3, are fine-tuned on
field photos of eight weed species and a Negatives class
[arxiv:1810.05726 § Deep Learning]. Each training image is resized to 256 × 256
and then rotated over the full circle, zoomed by a factor between 0.5 and 1,
sheared, shifted per colour channel by ±25 on the 8-bit scale, scaled in
brightness between 0.75 and 1.25, and flipped with probability 0.5. A centre
crop to 224 × 224 follows [arxiv:1810.05726 § Deep Learning]
[code: deepweeds.py#L174-L193]. Validation images go through the same chain
[code: deepweeds.py#L174-L193], so the validation loss that drives the learning
rate and early stopping is computed on randomly altered images
[code: deepweeds.py#L254-L256]. Card: [Olsen et al. (DeepWeeds)](/papers/1810.05726).

The paper describes random perspective transforms and a separate intensity
shift of ±25 on all channels. The released code has shear and a brightness
factor instead, with no perspective transform and no separate intensity shift
[arxiv:1810.05726 § Deep Learning] [code: deepweeds.py#L177-L182]. The text
does not say which chain produced the reported results.

The paper says that training without augmentation over-fitted strongly, but it
gives no numbers for that [arxiv:1810.05726 § Deep Learning]. The reported
class-weighted top-1 accuracy over the five test folds is 95.7% for ResNet-50
and 95.1% for Inception-v3 [arxiv:1810.05726 tab. 2]. No run without
augmentation is reported, so the effect of the chain alone is not measured.

### Random tone mapping for lighting-sensitive depth

This variant changes only the brightness of the RGB image. The depth map is
kept as it is, because a change of light does not move a surface. Each
training image gets a gamma curve with γ = 1/2.2. The image is then scaled so
that a chosen percentile of its intensity maps to 0.8, and the percentile is
drawn between 70 and 99 [arxiv:2408.06083 §3.1] [arxiv:2408.06083 eq. 1].
The paper states the scale only in words, and it does not say how the percentile
is sampled, whether it is drawn per sample or per iteration, or what intensity
range the input has [arxiv:2408.06083 §3.1]. Card:
[Zhang et al. (non-Lambertian depth)](/papers/2408.06083).

The motivation is that glass and mirror surfaces change in appearance under
light. Strong light can wash out glass texture, and weak light adds noise and
colour drift [arxiv:2408.06083 §1]. Applying the curve needs the original
scene assets, which the paper says exist for its synthetic training set
[arxiv:2408.06083 §3.1].

On the Booster region with ToM surfaces, δ1.05 rises from 0.644 for the
Depth Anything V2 baseline to 0.800 with this augmentation alone. The guidance
loss alone reaches 0.815, and the two together reach 0.859
[arxiv:2408.06083 tab. 3]. The baseline row of that ablation equals the public
model's row in the main table, and the paper does not say whether it was
retrained on the same data without this augmentation. The size of the
augmentation's own effect therefore rests on that unstated choice
[arxiv:2408.06083 §4.4.1]. The numbers are author-reported on the Booster
split with ground truth [arxiv:2408.06083 §4.1].

### Library transform pipeline with GPU batch augmentation

The fastai library builds a training pipeline from blocks. Per-item transforms
run on the CPU, and batch transforms run on the mini-batch, on the GPU when one
is available [arxiv:2002.04688 §2.1]. The Pets example uses a batch transform
set with a warp of zero, and it leaves rotation, zoom and lighting at their
defaults [arxiv:2002.04688 §2.1]. The paper gives no accuracy effect for
augmentation alone. Its defaults are a moving target: the paper describes
v2.0.0, and the defaults at the pinned code commit differ [arxiv:2002.04688 §1].
Card: [Howard & Gugger (fastai)](/papers/2002.04688).

### Tabular CutMix and mixup for contrastive pre-training

SAINT is a transformer over the cells of a table row. It can be pre-trained on
unlabelled rows, where each row has two views [arxiv:2106.01342 §4]. The second
view applies CutMix in raw input space: each cell may be replaced by the same
cell from a partner row, chosen by a binary mask [arxiv:2106.01342 eq. 3]. The
embeddings of that view are then mixed with the embedding of another row, with
weight α [arxiv:2106.01342 eq. 4]. An InfoNCE loss pulls the two views of a row
together, and a denoising loss rebuilds the original cells
[arxiv:2106.01342 §4] [arxiv:2106.01342 eq. 5]. Card:
[Somepalli et al. (SAINT)](/papers/2106.01342).

In the paper's ablation with 50 labelled rows, the mean AUC is 86.96 with both
augmentations, 85.37 with CutMix alone and 86.45 with mixup alone
[arxiv:2106.01342 tab. 9]. Pre-training gives 86.91 against 85.78 without it at
50 labels, and 92.99 against 93.13 when all labels are used
[arxiv:2106.01342 tab. 3]. The gain is therefore limited to few labels. All
figures are author-reported and averaged over the 14 binary datasets.

The released code differs from the paper in its defaults. It turns augmentation
off by default, uses 0.1 as the CutMix swap probability and 0.3 as the mixup
weight [code: train.py#L43] [code: train.py#L44] [code: train.py#L45]. The paper
gives p = 0.3 for the CutMix mask without saying whether it marks cells to keep
or to swap, so the swap rate is not settled by the paper
[arxiv:2106.01342 §5] [arxiv:2106.01342 eq. 3]. Predictions also depend on the
other rows in the batch, because intersample attention runs across the batch
[code: models/model.py#L121-L125].

### Resampling for class imbalance

A benchmark pairs nine resampling methods with nine ensemble models on
imbalanced tabular sets [arxiv:2304.02858 §4.1] [arxiv:2304.02858 tab. 1]. The
resamplers are SMOTE and four of its variants (SMOTE-ENN, Borderline-SMOTE,
SVM-SMOTE and K-means SMOTE), ADASYN, random over- and under-sampling, and
CT-GAN, which is a GAN [arxiv:2304.02858 tab. 1] [arxiv:2304.02858 §3.1]
[arxiv:2304.02858 §3.3]. The resampler acts on the training rows only, and the
classifier is unchanged [arxiv:2304.02858 §5.1]. Each pair is run 30 times on a
60:40 split [arxiv:2304.02858 §5.1]. Card:
[Khan et al. (imbalance benchmark)](/papers/2304.02858).

The conclusion names SMOTE with LightGBM and ROS with LightGBM as the best
pairs, and says that plain SMOTE and ROS beat CT-GAN at lower cost
[arxiv:2304.02858 §7]. The per-dataset best-pair table does not match that:
CT-GAN is the best pair on five of its rows [arxiv:2304.02858 tab. 8]. The
paper reports no timing or memory figures, so the cost claim has no data behind
it [arxiv:2304.02858 §7]. Read the tables rather than the conclusion.

Three points from the released code limit how the numbers can be read. The
loop resamples the output of the previous run, not the original training split,
so run-to-run variation differs from the protocol the paper describes
[code: raw/code/2304.02858/main.py#L218-L224]. The AUC column is computed from
hard predictions, which makes it balanced accuracy rather than a threshold-free
AUC [arxiv:2304.02858 eq. 21] [code: raw/code/2304.02858/main.py#L213]. The
split is not stratified, so a test set on a rare class can hold only a few
minority rows [arxiv:2304.02858 §5.1] [arxiv:2304.02858 tab. 2].

### Learned augmentation policies (survey only)

A survey of few-shot learning sorts augmentation into hand-crafted rules and
learned processing [arxiv:2205.06743 §3]. Hand-crafted rules include random
erasing, cutout-style filling, random cropping and mixing, and feature-level
variants [arxiv:2205.06743 §3.1]. Learned processing uses meta-learning to
choose or generate augmentations. The survey names an augmentation-policy search
(DADA), conditional generators (AFHN, MetaGAN) and a delta-encoder
[arxiv:2205.06743 §3.2]. It runs no experiment of its own, and its tables mix
protocols, so they cannot rank these methods against each other
[arxiv:2205.06743 tab. 9] [arxiv:2205.06743 tab. 12]. Card:
[Song et al. (few-shot survey)](/papers/2205.06743). The method is described
here for orientation only; read the original papers before implementing it.

### Text augmentation taxonomy (survey only)

A survey sorts text augmentation into paraphrasing, noising and sampling,
ordered by how much new diversity each adds [arxiv:2110.01852 §2]. Its
properties table marks Mixup as the only online method, and most non-learnable
methods as dependent on external resources such as WordNet [arxiv:2110.01852
§2.4] [tab. 1]. Its examples include EDA-style swaps, deletions and insertions
for noising [arxiv:2110.01852 §2.2], and word-piece replacement with
probability 0.4 [arxiv:2110.01852 §2.1.3]. The survey runs no experiments of
its own, so it gives no effect sizes [arxiv:2110.01852 §4]. Card: [Li, Hou & Che (NLP augmentation survey)](/papers/2110.01852).

### Input-side regularizers in a CNN survey (survey only)

A survey of recent CNN regularizers files Cutout, RandomErasing, AutoAugment,
PBA, RandAugment, Mixup and CutMix under input methods [arxiv:2201.03299 §1.3]
[arxiv:2201.03299 §3]. It reports that about 44% of the works it surveys change the input
[arxiv:2201.03299 §1.5]. It calls RandAugment the likely best input method,
which is the survey's opinion and was not tested [arxiv:2201.03299 §7.1]. The
Cutout patch side it lists is 16×16 for CIFAR-10 and 8×8 for CIFAR-100
[arxiv:2201.03299 §3.1]. Its compiled numbers are the original authors', not
reproduced [arxiv:2201.03299 tab. 2]. Card: [dos Santos & Papa (CNN regularization survey)](/papers/2201.03299).

### Manipulation, erasing and mixing measured across three tasks

A survey of image augmentation that also runs its own with and without
comparisons for the basic methods: manipulation, erasing and mixing
[arxiv:2204.08610 §4]. Printed changes are +0.53 to +2.61 mean IoU for four
segmenters on PASCAL VOC [arxiv:2204.08610 tab. 2], +0.07 to +2.80 accuracy for
three networks on CIFAR-10, CIFAR-100 and SVHN [arxiv:2204.08610 tab. 3], and,
on COCO2017, +0.40 mAP for Faster R-CNN but −0.27 for CenterNet
[arxiv:2204.08610 tab. 4]. Training settings are not stated and no code is
released [arxiv:2204.08610 §4]. Card: [Yang et al. (image augmentation survey)](/papers/2204.08610).

### Physics-rendered overlays on real photos

Paired training data for raindrop removal is hard to collect, so the authors
build it. They start from clean Cityscapes street photos and place virtual
raindrops on them. Each drop is a spherical cap in front of the camera. A ray
through a drop is refracted, and the pixel takes the background value where the
refracted ray meets the background plane. Drop regions are then blurred with a
disk kernel of 7 to 20 pixels. The drop positions come from the simulation, so
the binary drop mask comes with each image [arxiv:2609.37870 §3]. Each
background gets 50 to 70 drops. The training split has about 30,000 images from
the Cityscapes training set, and the test split has 1,525 rainy and clear pairs
from the Cityscapes test set [arxiv:2609.37870 §3] [arxiv:2609.37870 §5.1].

The clean photo is the target and the rendered photo is the input, so the label
stays valid by construction [arxiv:2609.37870 §3]. The base images are real and
labelled, which separates this from the 3D-scan case in the scope section below.
The rendering stage is under-specified. The sampling distributions, the
refraction equations and whether drop parameters are drawn per drop or per image
are not stated. No renderer or dataset release is mentioned
[arxiv:2609.37870 §3]. Card: [Hao et al. (raindrop removal)](/papers/2609.37870).

### Stitched face forgeries with known masks for deepfake detection

This variant makes forged training frames rather than transforming real ones.
A real target frame is paired with a frame of another identity whose 68 facial
landmarks are close to its own. The source face is aligned to the target, and
it is blended in through a mask. The mask starts as the convex hull of the
target landmarks, is deformed elastically, and is then blurred
[arxiv:2012.09311 §3.2] [arxiv:2012.09311 alg. 1]. The label is valid by
construction, because the method knows which pixels it pasted. Each real
sample in training is replaced by one of these forgeries with probability 0.5
[arxiv:2012.09311 §4.2]. The label is a pairwise consistency map and a
real/fake flag, not a class [arxiv:2012.09311 §3.1]. Card:
[Zhao et al. (self-consistency deepfake)](/papers/2012.09311).

The paper's ablation trains the same detector with and without these forgeries.
With DF frames as the real source, the four-set average AUC is 79.19 without
them and 86.37 with them. With DFDC-P as the source, it is 74.78 and 86.35
[arxiv:2012.09311 tab. 8]. The cross-dataset numbers train on real FF++ frames
plus these forgeries, and the in-dataset numbers use them by default, so the
two protocols are not directly comparable [arxiv:2012.09311 §4.2].

The blending method is picked at random from a set the paper does not list, and
the landmark distance threshold and the augmentation strengths are not stated
[arxiv:2012.09311 §3.2] [arxiv:2012.09311 alg. 1]. The released code is not
public, so these values have to be chosen again for a reimplementation. The
numbers are author-reported [arxiv:2012.09311 §4.3].

### Synthetic video from annotated still images

Labelled video for saliency is scarce, so the authors turn each annotated
still image into a training pair. They generate ten simulated frames per image
and use the warp I'(x) = I(x + v(x)), where v is a flow field
[arxiv:1702.00871 §4] [arxiv:1702.00871 eq. 5]. The field is built on
superpixels. Background parts are split into seeds that move at random and
the rest, and foreground parts share one main motion with small variation
[arxiv:1702.00871 §4]. A smoothing energy with a unary and a neighbour term
then keeps the seeds and the foreground near their initial motion
[arxiv:1702.00871 eq. 6] [arxiv:1702.00871 eq. 7] [arxiv:1702.00871 eq. 8].
The authors report more than 150K synthetic clips, generated in about three
hours [arxiv:1702.00871 §4] [arxiv:1702.00871 §5.1]. The mask is warped with
the same field, so the label moves with the pixels
[arxiv:1702.00871 §4].

The gain comes from mixing this data with real video. On the FBMS test set,
the full model reaches 7.65% MAE. Training on image-derived data alone gives
9.27%, and training on video alone, which is the smallest set, gives 24.5%
[arxiv:1702.00871 tab. 2] [arxiv:1702.00871 §5.3].

The paper does not state how pixels that leave the frame or become occluded
are filled, which interpolation the mask warp uses, or how many superpixels
are made [arxiv:1702.00871 §4]. The horizontal motion bound is also not stated
[arxiv:1702.00871 §4]. Reimplementers must choose these. Card:
[Wang et al. (video saliency)](/papers/1702.00871).

### Cross-painting robot demonstrations onto other arms

Robot demonstrations come mostly from a few arms, so a policy often works only
on the arm it saw. This variant reuses each recorded trajectory and renders it
with another robot. For each frame, a fused mask finds the source arm: a
learned segmenter, fine-tuned on a few labelled trajectories per dataset, and
a simulation mask, aligned to the image by a small grid search. A video
inpainting model fills the background behind the arm, and the target arm is
replayed in simulation along the same trajectory and composited over that
background [arxiv:2512.13100 §4.1] [arxiv:2512.13100 §4.2]. Each target base
position is tuned in simulation so that every pose is reachable, and a trajectory with no
feasible base is discarded [arxiv:2512.13100 §4.2 (2)]. Card:
[Ji et al. (OXE-AugE)](/papers/2512.13100).

The authors applied this to 16 Open X-Embodiment datasets, with eight or nine
target robots per dataset, and report more than 4.4M augmented trajectories
[arxiv:2512.13100 §6] [arxiv:2512.13100 app. A.2] [arxiv:2512.13100 tab. 1].
Two datasets get no WidowX augmentation, because most of their trajectories
exceed that arm's workspace [arxiv:2512.13100 app. A.2].

The authors report results from two setups. In simulation, with a Diffusion
Policy on five Robosuite tasks, adding one augmented target to the source
demonstrations improves success on that target by 9 to 19%, against the
augmented target data alone. Augmented training also makes the source robot
more robust under lighting and occlusion changes [arxiv:2512.13100 §5.1]
[arxiv:2512.13100 fig. 3] [arxiv:2512.13100 fig. 4]. In the real tests, OpenVLA-OFT and pi0 fine-tuned
with augmented Bridge data gain 24% and 45% on average, against fine-tuning
on Bridge alone, on unseen arm and gripper pairs [arxiv:2512.13100 §6.2]. The
text gives these gains without a unit or denominator, and the real-test score
is a partial credit, not a success rate [arxiv:2512.13100 app. A.5].

The paper also compares this with diffusion models trained to translate robot
appearance on 700K paired images. Fine-tuning on those images lowers final
policy performance by 27 to 30%, in the authors' setup [arxiv:2512.13100 §6.2]. The method assumes that
grippers are similar in shape and function, and that all frame transforms are
known [arxiv:2512.13100 §3]. Use it where those hold and where a simulator of
the target arm is available.

### Route-and-instruction pairs written by a multimodal model

Navigation agents that follow spoken directions learn from pairs of a route
and a sentence that describes it. This variant builds the pairs from ordinary
house-tour videos instead of a simulator. Each video is sampled as a path of
room nodes and transition nodes. CLIP labels each room with its type and
salient objects, and an action inverse model estimates the action between
consecutive rooms [arxiv:2411.11394 §3.1]. A multimodal model then receives the
room images, labels and actions in order, and writes an instruction for the
path [arxiv:2411.11394 §3.2]. Card: [Yan et al. (InstruGen)](/papers/2411.11394).

The check is what keeps the label valid. The model is asked to extract the
room and action pairs from its own instruction. If they differ from the route's
labels, the instruction is regenerated, and regular expressions remove stray
characters and prompt fragments [arxiv:2411.11394 §3.2]. The check compares the
instruction with the route's labels, so an error in CLIP or in the action
model passes through it unless it breaks the instruction's logic
[arxiv:2411.11394 §3.1] [arxiv:2411.11394 §3.2]. The mismatch test, the retry
limit and the number of accepted pairs are not stated.

The ablation on R2R val-unseen adds the steps one at a time. Without any
check, the success rate is 67.42. Adding the consistency check gives 69.32, the
regex step 70.90, and a manual review of the best row 71.10 [arxiv:2411.11394 tab. 1].
The last step is the only one that is not automatic, and the paper does not
say how much review was done. The automatic-only row is therefore the figure to
compare when no one can review the pairs.

The prompt can ask for a concise or a detailed instruction [arxiv:2411.11394 §3.2].
In the granularity table, coarse-grained pretraining beats fine-grained
pretraining for every fine-tuning set. The best row pairs coarse-grained
pretraining with the R2R* fine-tuning set, at 71.10 success on R2R val-unseen
[arxiv:2411.11394 tab. 2]. The card does not establish how the coarse-grained
set maps onto the concise and detailed prompts. R2R* mixes the generated instructions
with the original R2R ones, and the mix ratio is not stated [arxiv:2411.11394 §4.2].
The paper's comparison table places this above Lily, at 70.00
[arxiv:2411.11394 tab. 3]. The one-shot result, with one training environment
averaged over five random sets, is 51.25 against 50.86 for Lily, a margin under
one point with no spread reported [arxiv:2411.11394 tab. 4]. The RxR zero-shot
score is 29.35, with a model trained on R2R* only [arxiv:2411.11394 tab. 5].
The comparison table does not say whether its baseline rows were re-run or
copied [arxiv:2411.11394 tab. 3].

The released repository holds only a README, so the prompts, the regex rules
and the generated set cannot be checked. The prompt appears only as a figure,
and the learning rate, the optimizer and the pretext-loss weights are not
stated [arxiv:2411.11394 fig. 2] [arxiv:2411.11394 §4.1]. The generated set's
size is not given beyond the 3,879 source videos [arxiv:2411.11394 §3.1]. The
model's version, temperature and sampling settings are not stated
[arxiv:2411.11394 §3.2].

### Laser-beam band swap between scans for semi-supervised LiDAR segmentation

Each scan is split into areas by laser inclination. Two scans then exchange
alternating areas, so that every area's neighbours come from the other scan
[arxiv:2405.05258 §3.2] [eq. 7]. A teacher, whose weights are an exponential
moving average of the student's, pseudo-labels the unlabelled scans, and only
predictions above a confidence threshold are kept [arxiv:2405.05258 §3.3].
LaserMix++ adds camera features: each point is paired with its image pixel,
the number of areas is drawn from 2 to 6, and point features are distilled
towards image features [arxiv:2405.05258 §4.2] [arxiv:2405.05258 §4.3]. Card:
[Kong et al. (LaserMix++)](/papers/2405.05258). The authors report an average
gain of 2% to 3% for LaserMix++ over LaserMix [arxiv:2405.05258 §5.3], and a
gain of 3.1 to 5.6 mIoU with Cylinder3D at 1% labels [arxiv:2405.05258 tab. 2].
The linked repository holds the LaserMix code only, so the camera parts are
not reproducible from it [code: README.md#L177], and several released config
values differ from the paper, including the pseudo-label threshold.

### Object-first LiDAR generation for detector augmentation

Real LiDAR objects are sparse and partial, and foreground points are a small
share of a scene's points. The authors say that a scene-level generator
therefore spends most of its capacity on background [arxiv:2412.17226 §3.2].
This variant generates each foreground object first, from a text description
and a 3D box [arxiv:2412.17226 §3.2]. Each object comes from a point-cloud
diffusion model that takes a CLIP text embedding and a Fourier embedding of
the box as its condition [arxiv:2412.17226 eq. 1] [arxiv:2412.17226 eq. 2]
[arxiv:2412.17226 eq. 3]. The objects are then projected into a range image,
and a scene range-image diffusion model receives them through a
ControlNet-style controller [arxiv:2412.17226 §3.2] [arxiv:2412.17226 eq. 7].
Card: [Yan et al. (OLiDM)](/papers/2412.17226).

An extra per-category loss, called object semantic alignment, splits the scene
range image into one channel per category and adds a second noise-prediction
loss for it [arxiv:2412.17226 §3.3] [arxiv:2412.17226 eq. 9]. The paper adds
this to the scene loss, but it does not state the weighting
[arxiv:2412.17226 §3.3]. The method assumes that each training object has a
category, a caption and a 3D box [arxiv:2412.17226 §3.2] [arxiv:2412.17226 app. A].
Captions come from BLIP-2 [arxiv:2412.17226 app. A].

Scene-level generation is compared on KITTI-360. Against UltraLiDAR, the
paper reports FPD 7.60 against 25.1 and JSD 0.42 against 0.71
[arxiv:2412.17226 tab. 1]. Removing the per-category loss raises FPD to 9.20
[arxiv:2412.17226 tab. 1], so the paper credits that loss with part of the
gain. Foreground quality is reported in a separate table, where the full model
has the lowest Chamfer distance, 1.88, against 17.6 for R2DM
[arxiv:2412.17226 tab. 2].

The downstream claim is about detection. Generated objects are placed on the
ground plane, uniformly, to avoid oversampling near the ego vehicle
[arxiv:2412.17226 app. E]. On nuScenes, PointPillars goes from 44.8 mAP with
no augmentation to 46.7 with OLiDM objects, against 45.1 with ground-truth
copy-paste augmentation. CenterPoint goes from 59.0 to 61.9
[arxiv:2412.17226 tab. 4]. The number of objects added is not stated
[arxiv:2412.17226 app. E]. The paper reports no standard deviation or seed
[arxiv:2412.17226 §4.1].

Training costs are reported in the appendix. OLiDM takes 19 hours of training
on the authors' KITTI-360 setup, against 12 hours for R2DM, and 3.5 hours of
inference [arxiv:2412.17226 app. C]. The cost table gives no hardware and no
sample count for the inference figure. Scene generation runs 300k steps at
batch 16 [arxiv:2412.17226 app. A].

Eq. 5 as printed does not define an attention operation on its own, and eq. 7
and eq. 8 disagree on what the controller receives. The paper does not say
which version was trained [arxiv:2412.17226 eq. 5] [arxiv:2412.17226 eq. 7]
[arxiv:2412.17226 eq. 8]. Plan the object denoiser's attention yourself, and
check the released code before reproducing the numbers. The released repository
has no model code, so there is nothing to check against yet.

### 3D rotation, scaling and shear for skeleton sequences

Skeleton action recognition takes 3D joint coordinates as input, so image
transforms do not apply directly. The two-stream RNN perturbs those
coordinates during training only [arxiv:1704.02581 §4.3]. Three transforms
are used. The rotation is composed from three axis rotations, and the input
is rotated at random "within a certain range" about the x and y axes. The
authors tie this to viewpoint change, with the camera plane perpendicular to z
[arxiv:1704.02581 §4.3] [arxiv:1704.02581 eq. 5]. Scaling is a diagonal matrix
with scale factors for each axis, drawn per sample to cover different body
sizes [arxiv:1704.02581 eq. 6] [arxiv:1704.02581 §4.3]. Shear is a 3×3 matrix
with six off-diagonal factors [arxiv:1704.02581 eq. 7]. Card:
[Wang & Wang (two-stream skeleton RNN)](/papers/1704.02581).

The paper gives no ranges, probabilities or copy counts for any of the three
[arxiv:1704.02581 §4.3]. The text applies the rotation to x and y, but the
composite in eq. 5 has three axes, and whether a z rotation is drawn is not
stated [arxiv:1704.02581 eq. 5]. The card's search found no official code, so
an implementer has to choose the ranges.

In the authors' NTU ablation, the two-stream model without a transform reaches
68.6% cross-subject and 71.7% cross-view. With the 3D transform it reaches
71.3% and 79.5% [arxiv:1704.02581 tab. 1]. On SBU the same change moves 91.9%
to 94.8% [arxiv:1704.02581 tab. 1]. The table row does not say which of the
three transforms it includes, so the ablation cannot separate them
[arxiv:1704.02581 tab. 1]. Use this where a change of camera angle or body
size should leave the action label unchanged.

## When to use which

| Situation | Use | Why |
|---|---|---|
| Instance masks exist, Mask R-CNN-style detector, few labels | Copy-Paste compositing | Largest reported gain with scarce labels; needs masks, not only boxes [arxiv:2012.07177 §4.3] [arxiv:2012.07177 §3] |
| Classifier with a CutMix-style mixed-label recipe, no masks, and extra saliency compute per mixed batch is acceptable | Saliency-guided patch mixing | Gains over CutMix are clear on CIFAR without crop and flip, and small on ImageNet ResNet-50 [arxiv:2006.01791 tab. 1] [arxiv:2006.01791 tab. 2] |
| Small domain with a fixed layout and known symbol invariances | Hand-designed label-preserving transforms | Only valid when the invariance holds for your symbols; the authors chose the valid flips per dataset [arxiv:1802.09567 tab. 3] [arxiv:1802.09567 §5.1.4] |
| Small labelled image set with no known domain invariance | Standard geometric and photometric transforms | Measured with and without on one network for histopathology; at 40x the binary gain is 0.8 points and the eight-class gain 1.4 points [arxiv:1811.04241 tab. 3] [arxiv:1811.04241 tab. 4]. The face-mask paper names its transforms but records no comparison [arxiv:2009.08369 §4.1] |
| Pixel-wise segmentation of large images that cannot be resized, rare classes, texture with no preferred orientation | Rotated patches with class-balanced strides | Rotation on top of balancing raises SEM pixel accuracy from 92.11% to 93.92% [arxiv:1706.06480 tab. 5]; the rotation's label validity is assumed, not tested [arxiv:1706.06480 §6] |
| Few labels per class, frozen pretrained embedding, non-negative features | Feature-space sampling | No backbone training; 1-shot miniImageNet goes from 56.37 to 68.57 with both steps [arxiv:2101.06395 tab. 4] |
| One satellite or reference image per class, batch size known, extra training time affordable | Repeated augmented copies | Copies from 1 to 3 add 1.9 AP in the drone-to-satellite direction; inference cost is unchanged [arxiv:2201.09206 tab. 7] [arxiv:2201.09206 §3.3] |
| Graph data with no labels, pre-training for a downstream node task | Graph augmentations in contrastive pre-training | Views come from the graph itself; the augmentation choice needs empirical tuning [arxiv:2103.00111 §3.2] |
| Unlabelled time series with a labelled downstream task, and a frozen-encoder probe for evaluation | Weak and strong views in contrastive pretraining | Linear probe reaches 90.37% on HAR against 90.14% supervised, and 83.00% on Sleep-EDF against 83.41% [arxiv:2106.14112 tab. 2]; single-view ablations hurt HAR most [arxiv:2106.14112 tab. 4] |
| Labelled entity sentences, a large pool of unlabelled abstracts, and a tagger to score the candidates | Model-filtered sentences | Gain of 0.51 F1 in one run on a 311-sentence test split [arxiv:2606.29836 tab. 7]; the procedure has no released script [arxiv:2606.29836 §3.3] |
| Training a ViT or hybrid from scratch on small data | Strong augmentation bundle | Recipe names the components; values and their individual effects are not stated [arxiv:2202.06709 app. A] [arxiv:2202.06709 app. F] |
| Labelled sentence pairs, large unlabelled pool, a cross-encoder you can run at training time, and a fast bi-encoder for serving | Recombined pairs with teacher labels | In-domain gains of about 1.0 to 6.0 points; random pairs hurt on Spanish-STS [arxiv:2010.08240 tab. 5]. Each candidate needs a cross-encoder pass [arxiv:2010.08240 §3.1] |
| Labelled source domain, unlabelled target-domain pairs, no target labels | Recombined pairs with teacher labels (domain adaptation) | Quora to Sprint rises from 0.505 to 0.875; the Bi-LSTM baselines of the paper are level or better on Sprint as target [arxiv:2010.08240 tab. 6] |
| Class-labelled images, a text-to-image generator you can fine-tune, and a large sampling budget | Generated images | +1.78 top-1 for ResNet-50 at 1.2M added images [arxiv:2304.08466 tab. 3]; the gain is gone by about 6M total [arxiv:2304.08466 tab. 4]. Fine-tuning needs 256 TPU-v4 chips [arxiv:2304.08466 §4.1] |
| Synthetic data with known ground truth, and a reasoning model that reads a graph predicted by an earlier stage | Graph perturbations | On TGQA with Llama2-13B, exact match rises from 0.753 to 0.797 with the four perturbations; single runs [arxiv:2401.06853 tab. 5]. Fix the seed, since the code draws the entity map once per run [code: src/SFT_TG_Reasoning.py#L101-L105] |
| Detector training where a recipe names mosaic and mixup but gives no parameters | Mosaic and mixup, measured on your data | The paper gives no effect size or parameters [arxiv:2408.15857 §3.2] [arxiv:2408.15857 §8] |
| A small photo set, rotations and flips keep the label, split not yet made | Flips and rotations, split first | Grew 4,500 to 27,000 images; the split order is not stated [arxiv:2609.10524 §4.1] [arxiv:2609.10524 tab. 2]. Splitting before augmenting keeps copies of one photo on one side (our recommendation) |
| Tabular minority class with several distinct failure types, a GAN you can fit per type, and many hours of compute | Per-type GAN oversampling | PR-AUC 0.8574 against 0.835 for the cost-sensitive forest, one run per configuration [arxiv:2607.19153 tab. 4] [arxiv:2607.19153 §4.1]; the single GAN and SMOTE rows score lower in the same table [arxiv:2607.19153 tab. 4]. About 14 h of training [arxiv:2607.19153 §5.1] |
| Forecasting with sharp changes between readings, where a doubled input window and longer training are affordable, and you can define the virtual values yourself | Virtual values between samples | One-step MAPE 0.98–0.99% against 1.26–1.31% for the same model without them, on one data set [arxiv:2512.06714 tab. 6]; the generation rule is not stated [arxiv:2512.06714 §3.5] |
| Labelled text is scarce, a capable language model can write candidates, and you will filter them yourself | Language-model-written examples | The survey describes filtering of noisy generated examples but gives no result for it [arxiv:2111.01243 §5.1] [arxiv:2111.01243 §5] |
| Small set of field photos, ImageNet-pretrained CNN, invariances unknown | Heavy transform chain (rotation, zoom, shear, colour, flip) | Reported 95.7% weighted top-1 for ResNet-50 with no run without augmentation [arxiv:1810.05726 tab. 2] [arxiv:1810.05726 § Deep Learning]; the paper and code chains differ [code: deepweeds.py#L174-L182] |
| Depth model trained on synthetic indoor scenes, with glass and mirror surfaces that change appearance under light, and the original scene assets to hand | Random tone mapping | Booster ToM δ1.05 from 0.644 to 0.800 with this augmentation alone, and 0.859 with the guidance loss too [arxiv:2408.06083 tab. 3]; the baseline row's training is not stated [arxiv:2408.06083 §4.4.1] |
| Unlabelled table rows for pre-training, few labels, row-attention transformer | CutMix and mixup in contrastive pre-training | 86.91 against 85.78 without pre-training at 50 labels; 92.99 against 93.13 with all labels [arxiv:2106.01342 tab. 3]; each augmentation alone trails both [arxiv:2106.01342 tab. 9] |
| Tabular data with a rare class, a gradient-boosting or forest classifier, time to rerun a benchmark | Oversampling with SMOTE or ROS before the classifier | The conclusion names SMOTE and ROS with LightGBM as best [arxiv:2304.02858 §7]; its best-pair table names CT-GAN best on five rows [arxiv:2304.02858 tab. 8]; read the table, not the summary |
| Contrastive pre-training with strong views, evaluated by a linear probe | Check for dimensional collapse before trusting strong views | Collapse appears in SimCLR; the theory that ties it to strong augmentation covers only linear networks [arxiv:2110.09348 §4.2] [arxiv:2110.09348 §6.2]; augmentation probabilities are in the code, not the paper [arxiv:2110.09348 app. D] |
| Restoration task with clean and degraded image pairs that are hard to collect, real clean photos to overlay, and a renderer you can rebuild from the text | Physics-rendered overlays on real photos | Author-reported 41.29 dB PSNR on their synthetic test set against 37.79 dB for Qian et al. [arxiv:2609.37870 tab. 1]; on the real-image set of Qian et al. the rough-mask model reaches 30.17 dB against 30.82 dB for Qian et al. [arxiv:2609.37870 tab. 2]. No renderer or dataset release is mentioned [arxiv:2609.37870 §3] |
| Binary forgery detector with real face frames, no manual manipulation masks, and a consistency head you can train | Stitched face forgeries with known masks | Four-set average AUC 79.19 without the forgeries and 86.37 with them, DF training [arxiv:2012.09311 tab. 8]; the blending set and landmark threshold are not stated [arxiv:2012.09311 §3.2] |
| Attributed graph with no labels, pre-training for node classification, and centrality scores you can compute once | Centrality-adaptive drops | Adaptive beats uniform drops by 0.1 to 1.6 points on five datasets, largest on Amazon-Computers [arxiv:2010.14945 tab. 4]; the reported gains are in-paper and the official code differs in details [arxiv:2010.14945 §3.3.1] |
| Graph contrastive pre-training where you cannot justify a fixed edge-drop rule, and a second network per minibatch is affordable | Learned edge drops | Author-reported significant gains over InfoGraph and GraphCL on 5 to 6 of 8 TU datasets [arxiv:2106.05819 §5.1]; the drop ratio is set by λ. Each minibatch runs the encoder again for the augmenter update; the paper's COLLAB timing is 14.4 minutes against 3.2 hours for JOAO [arxiv:2106.05819 app. H] |
| Video task with annotated still images but little or no labelled video, and a warp you can reimplement | Synthetic video from still images | On FBMS, image-derived data alone gives 9.27% MAE and video alone 24.5%, against 7.65% with both [arxiv:1702.00871 tab. 2] |
| Closed-set classifier also used to reject unknown classes, with hundreds of epochs affordable | RandAugment and label smoothing recipe | TinyImageNet AUROC: +1.8 from RandAugment at 200 epochs and +4.6 from label smoothing at 600 [arxiv:2110.06207 app. C]; single-setting ablations |
| Real-time one-stage detector trained from scratch, with image loading as a bottleneck | Cached Mosaic and MixUp, then large-scale jittering | RTMDet-l reaches 51.3 AP against 46.7 with large-scale jittering alone [arxiv:2212.07784 tab. 7]; cached Mosaic is 3.6 times faster per image in the authors' timing [arxiv:2212.07784 tab. 7] |
| Robot demonstrations from few arms, a simulator model of each target arm, similar grippers, and a segmentation model you can fine-tune | Cross-painting demonstrations onto other arms | Author-reported +24% (OpenVLA-OFT) and +45% (pi0) on unseen arm and gripper pairs, with no unit given [arxiv:2512.13100 §6.2]; diffusion appearance translation lowers success by 27 to 30% in the same paper [arxiv:2512.13100 §6.2] |
| Vision-and-language navigation with house-tour videos and no simulator scenes, and a multimodal model you can call with a consistency check | Route-and-instruction pairs from a multimodal model | Author-reported 71.10 SR on R2R val-unseen, against 70.00 for Lily; the automatic-only row gives 70.90 [arxiv:2411.11394 tab. 1] [arxiv:2411.11394 tab. 3]; the prompt appears only as a figure [arxiv:2411.11394 fig. 2] |
| LiDAR detection where foreground objects are rare in scenes, and a diffusion model you can train on object boxes and categories | Object-first LiDAR generation | Author-reported mAP 46.7 for PointPillars with generated objects against 44.8 with none, on nuScenes [arxiv:2412.17226 tab. 4]; the paper reports no seeds or standard deviations, and the released repository has no model code |
| Image classifier in the fastai library, GPU available, and the defaults are acceptable | Library batch transforms | The paper gives no accuracy effect of augmentation; the Pets example uses a warp of zero and default rotation, zoom and lighting [arxiv:2002.04688 §2.1] |
| Skeleton (3D joint) sequences where camera angle and body size vary, and you can choose the transform ranges | 3D rotation, scaling and shear | NTU cross-view rises from 71.7% to 79.5% and cross-subject from 68.6% to 71.3% with the transform, author-reported [arxiv:1704.02581 tab. 1]; ranges not stated [arxiv:1704.02581 §4.3] |
| A few hundred colour-varied images, a pretrained frozen encoder you can run many times per image, and a tree classifier on top | Colour and crop views pooled over frozen features | Author-reported 87.2% four-class accuracy under 10-fold cross-validation [arxiv:1802.00752 tab. 1]; the per-fold spread is 2.6 points [arxiv:1802.00752 tab. 1] |
| Contrastive pretraining with a MoCo-style queue, and a wall-clock budget about a third larger than the baseline | Strong views matched through similarity distributions | 69.4 against 67.5 for MoCo v2 at 200 epochs [arxiv:2104.07713 tab. 2]; the strong view alone scores 67.7 [arxiv:2104.07713 tab. 5]; the cost is 70 h against 53 h [arxiv:2104.07713 tab. 6] |
| Driving LiDAR segmentation with few dense labels, a teacher-student setup, and a LiDAR-only pipeline (use LaserMix++ only if the camera parts are reproducible) | Laser-beam band swap between scans | Author-reported gain of 2% to 3% for LaserMix++ over LaserMix [arxiv:2405.05258 §5.3]; the released code omits the camera parts [code: README.md#L177] |
| Speech with a few hours of labels, a pretrained speech encoder, and compute for several retrainings | Noisy-student loop with audio augmentation | WER falls from 0.568 to 0.515 over five generations [arxiv:2609.24410 tab. 4]; augmentation magnitudes are not stated [arxiv:2609.24410 §6.2] |
| Text task where you need an overview of augmentation families before choosing a method | Text augmentation taxonomy | Survey with no experiments of its own [arxiv:2110.01852 §4]; Mixup is the only online method in its table [arxiv:2110.01852 tab. 1] |
| CNN classifier with a fixed training budget, and a map of input, internal and label regularizers | Input-side regularizers from a CNN survey | The survey's compiled numbers are the original papers' and are not reproduced [arxiv:2201.03299 tab. 2]; its claim that RandAugment is likely best is an opinion [arxiv:2201.03299 §7.1] |
| Segmentation, classification or detection benchmarks, and a fixed recipe you can run with and without augmentation | Manipulation, erasing and mixing methods | Printed gains of +0.53 to +2.61 mean IoU for segmentation and +0.07 to +2.80 accuracy for classification [arxiv:2204.08610 tab. 2] [arxiv:2204.08610 tab. 3]; detection gains only for Faster R-CNN, and CenterNet loses 0.27 mAP [arxiv:2204.08610 tab. 4] |
| Plenty of labelled data | Expect smaller gains | Copy-Paste gains shrink with longer schedules and more data [arxiv:2012.07177 §4.2] [arxiv:2012.07177 §4.3] |
| Random pairing would paste objects into implausible scenes | Scene-aware pasting, not evaluated here | The paper's random baseline works, but it mixes scene types often [arxiv:2012.07177 app. C]; the context-based alternatives are not compared in these papers |

## Implementation sketch

The Copy-Paste step from [arxiv:2012.07177 §3] in Python-flavoured
pseudocode. Both images are assumed to be resized to one canvas size before
compositing; the paper pads small images with gray [arxiv:2012.07177 §3].

```pseudocode
# Copy-Paste for one training example  [arxiv:2012.07177 §3]
# Input:  labelled set D with masks, boxes and classes per instance
# Output: composite image and its updated instance list

def copy_paste_example(D):
    # Sample two training images at random  [arxiv:2012.07177 §3]
    main, paste = sample(D, 2)
    # Scale-jitter and flip each image on its own  [arxiv:2012.07177 §3]
    main = scale_jitter_and_flip(main)
    paste = scale_jitter_and_flip(paste)
    # Choose a random subset of the pasted image's objects  [arxiv:2012.07177 tab. 10]
    chosen = random_subset(paste.instances)
    # Union of the chosen masks, binary, shape (H, W)  [arxiv:2012.07177 §3]
    alpha = union([inst.mask for inst in chosen])
    # Composite the pasted pixels over the main image  [arxiv:2012.07177 §3]
    image = paste.image * alpha + main.image * (1 - alpha)
    # Trim or drop main-image objects covered by the paste  [arxiv:2012.07177 §3]
    kept = []
    for inst in main.instances:
        mask = inst.mask * (1 - alpha)
        if mask.sum() == 0:
            # Fully covered: remove from the labels
            continue
        # Partly covered: keep, with box recomputed from the new mask
        kept.append(inst.with_mask(mask))
    return image, kept + chosen
```

Where the variants differ: the graph variant replaces the composite step with
feature masking or edge edits applied to two copies of a graph. The
label-preserving variants replace it with a fixed set of flips, colour
inversions or generic geometric transforms [arxiv:2103.00111 §6.1]
[arxiv:1802.09567 §4.2] [arxiv:1811.04241 §4]. The saliency variant keeps the
paste step but chooses the patch from a saliency peak and mixes labels by patch
area [arxiv:2006.01791 §3.2]. The feature-space variant
skips images entirely and samples vectors from a Gaussian per support feature
[arxiv:2101.06395 §3.2.3]. The repeated-copy variant keeps the image pipeline
and changes only how many times each scarce example is drawn per step
[arxiv:2201.09206 §3.3]. The sentence-pair variant samples pairs instead of
transforming them, and the generative variant replaces the input image with a
class-name prompt [arxiv:2010.08240 §3.1] [arxiv:2304.08466 §4.1]. The
temporal-graph variant changes graph facts, relation names, entity names and
years together, using the same mapping for every field [arxiv:2401.06853 app. B].
The time-series variant makes two views of one series, one weak and one strong,
and the model-filtered variant samples new sentences that a tagger has checked
[arxiv:2106.14112 §3.1] [arxiv:2606.29836 §3.3]. The per-type GAN variant
replaces the transform with one generator per failure type and pools its
filtered samples [arxiv:2607.19153 §3.2.5]. The virtual-value variant inserts
a value between each pair of readings, with the rule left to the implementer
[arxiv:2512.06714 §3.5]. The heavy-chain variant is the standard per-image
pipeline with a fixed order of steps [code: deepweeds.py#L174-L182]. The tabular
variant swaps cells between rows and mixes embeddings [arxiv:2106.01342 §4].
The resampling variant adds minority rows before fitting and leaves the
classifier unchanged [arxiv:2304.02858 §4.1]. The learned edge-drop variant
replaces the fixed drop rule with a probability per edge, produced by a
network that is trained adversarially against the encoder [arxiv:2106.05819
§3.2]. The synthetic-video variant warps a still image along a flow field to
make the second frame of a training pair [arxiv:1702.00871 §4]. The robot
cross-painting variant replaces the image transform with a segment, inpaint and
replay pipeline, and drops trajectories with no feasible target base
[arxiv:2512.13100 §4.1] [arxiv:2512.13100 §4.2]. The route-instruction variant
replaces the image step with a path sampler, a labeller and a checked text
generator [arxiv:2411.11394 §3.1] [arxiv:2411.11394 §3.2]. The LiDAR variant
generates objects first and then places them in a range-image scene model
[arxiv:2412.17226 §3.2]. The skeleton variant perturbs the joint coordinates
with a random 3D rotation, scaling and shear before either stream reads them
[arxiv:1704.02581 §4.3].

## Contradictions

- **Random patch placement.** Copy-Paste pastes objects at random positions and
  reports gains [arxiv:2012.07177 §3]. SaliencyMix argues that a random patch can
  land on background and teach the model that background belongs to the source
  class [arxiv:2006.01791 §1]. The tasks differ (instance segmentation against
  classification), so the two claims are not directly comparable.
- **SaliencyMix claims against its tables.** The text calls its CIFAR-10 result of
  2.76% the best known top-1 error, and says it beats all other methods except
  PuzzleMix on CIFAR [arxiv:2006.01791 §4.1.1]. Table 1 has AutoAugment at 2.60%
  on that column [arxiv:2006.01791 tab. 1]. On ImageNet, the text says SaliencyMix
  beats all other methods, but table 2 has PuzzleMix ahead on top-1 (21.24
  against 21.26) and top-5 (5.71 against 5.76) [arxiv:2006.01791 §4.1.2]
  [arxiv:2006.01791 tab. 2]. Use the tables, not the text.
- **Source of the ALPR gain.** The card for [arxiv:1802.09567] reports that
  augmentation is essential for letter recognition, while the paper's own
  discussion attributes the main gain to splitting letters and digits into two
  networks, with no ablation separating the two [arxiv:1802.09567 §5.1.4].
  Treat the size of the augmentation effect as unmeasured.
- **Graph augmentation numbers.** The survey's accuracy table combines results
  from different papers and protocols, and it does not say which values were
  re-run [arxiv:2103.00111 §8]. Gaps between methods on that table are not a
  controlled comparison of augmentations.
- **Which baseline row is the k = 1 model (FSRA).** The k = 1 row of the main
  comparison matches the row with both the KL loss and the triplet loss in the
  ablation, while the text says the KL loss was not used in the final model
  [arxiv:2201.09206 tab. 2] [arxiv:2201.09206 tab. 7] [arxiv:2201.09206 §3.4].
  The triplet-only row differs by 0.03 AP, so the size of the copy effect is
  not affected, but the k = 1 baseline is ambiguous.
- **Fixed against learned graph drops.** Both the centrality-based method and
  the learned method report gains over uniform or non-adversarial drops, but
  on different tasks and protocols. The centrality method is evaluated on five
  node classification datasets [arxiv:2010.14945 tab. 4]. The learned method
  is evaluated on graph classification with linear probes. Its paper says an
  SVM protocol makes untrained encoders look much better, which is why it
  avoids that protocol [arxiv:2106.05819 §5.1]. The two sets of numbers cannot be ranked against
  each other.
- **Multi-GAN F1 in the text against its table.** The text calls the multi-GAN
  F1 of 0.6923 the highest of all conditions, but the baseline forest's F1 in
  [tab. 4] is 0.740 [arxiv:2607.19153 §4.1] [arxiv:2607.19153 tab. 4]. Use the
  table, not the sentence.
- **CT-GAN in the imbalance benchmark.** The conclusion says CT-GAN is not a top
  performer [arxiv:2304.02858 §7]. The best-pair table names CT-GAN best on five
  rows [arxiv:2304.02858 tab. 8]. Both cannot hold; use the table.
- **Raindrop removal claims against its tables.** The introduction claims a
  clear lead over every existing method on all existing datasets and on real
  rainy images [arxiv:2609.37870 §1]. On the real-image set of Qian et al.,
  Table 2 has Qian et al. ahead on PSNR, 30.82 dB against 30.17 dB for the
  rough-mask model. The rough-mask model has the higher SSIM, 0.9128 against
  0.9050 [arxiv:2609.37870 tab. 2]. The text calls the result "on par"
  [arxiv:2609.37870 §5.2]. Use the table, not the introduction.
- **Cached-Mosaic gain for small models.** The text says a small FIFO cache
  helps the small models by about 0.5 AP [arxiv:2212.07784 §4.4]. Table 7 gives
  44.2 against 43.9 for RTMDet-s, a gain of 0.3, and 51.1 against 51.3 for
  RTMDet-l, a loss of 0.2 [arxiv:2212.07784 tab. 7]. Use the table.
- **Mixing count in RTMDet.** The text says eight images are mixed per sample
  [arxiv:2212.07784 §3.3]. The official configs combine four images in Mosaic
  and two in MixUp [code: rtmdet_l_8xb32-300e_coco.py#L67-L81]. Treat the
  count as the config's, not the text's.
- **Adaptive gain on Amazon-Computers.** The text says the full adaptive scheme
  gives 1.5 points over the base model [arxiv:2010.14945 §4.3]. Table 4 gives
  87.85 against 86.25, a difference of 1.60 [arxiv:2010.14945 tab. 4].
- **Printed changes against printed values (image survey).** The "reported
  change" for DeepLabV3+ is +0.53, but its printed values 75.32 and 75.75 differ
  by 0.43 [arxiv:2204.08610 tab. 2]. The SVHN Wide-ResNet row prints +0.80, but
  its values differ by 0.08 [arxiv:2204.08610 tab. 3]. For Faster R-CNN, the
  mAP change printed as 2.40 is 0.40 from the printed values [arxiv:2204.08610
  tab. 4]. Use the values, not the printed change column.
- **Relative WER gain (speech).** The text calls the gain an "8% relative
  improvement", but the card's computation from Table 4 gives 9.3%
  [arxiv:2609.24410 §9] [arxiv:2609.24410 tab. 4].
- **Detection gains in the OLiDM paper against its table.** The headline gives
  mAP +2.4 and NDS +1.9 on nuScenes [arxiv:2412.17226 abstract]. The mAP figure
  is the mean of the two mAP gains in the table, but the NDS gains in
  [tab. 4] are +2.8 for PointPillars and +2.7 for CenterPoint, so +1.9 does not
  match. The table also labels the GT-Aug row "Real" for PointPillars and
  "Synthetic" for CenterPoint [arxiv:2412.17226 tab. 4]. Use the table, and
  treat the grouping of GT-Aug as unresolved.

## Scope: what is not augmentation

Some steps in the cited papers add or select training data without
transforming existing examples. They are not variants on this page.

Omnidata renders new views and labels from textured 3D scans, so it creates
labelled examples rather than transforming existing ones [arxiv:2110.04994 §3.1].
It is a dataset-generation method, and it is not one of the variants above; see
the `dataset-curation` concept for dataset construction. Card: [Eftekhar et al. (Omnidata)](/papers/2110.04994).
The released starter data and weights are licensed for non-commercial research
only, which matters for any use of its outputs; the card's Pitfalls section
has the details.

Test-time augmentation is also out of scope for training. It averages a
classifier's outputs over transformed copies of one input, and the weights for
combining them are learned on held-out data [arxiv:2011.11156 §4]. The paper
reports that plain averaging can turn correct predictions wrong
[arxiv:2011.11156 §3.2]. Card: [Shanmugam et al. (TTA aggregation)](/papers/2011.11156).

Random-view scoring for computer-aided detection is also test-time averaging.
Each candidate is cut into random translated and rotated cubes at several
scales, and the ConvNet's probabilities are averaged over them
[arxiv:1505.03046 §II-C] [arxiv:1505.03046 eq. 1]. The aggregation is applied
to the views at scoring time, so it belongs with the test-time entry above.
Card: [Roth et al. (random view aggregation)](/papers/1505.03046).

Some cards name augmentation without describing it, so they give nothing to
build a variant from. The drug-discovery review proposes synthetic data for
scarce data but describes no method and reports no experiment
[arxiv:2212.08104 §7]. The emotion-detection paper says that data augmentation,
oversampling and re-weighting are used, but states no factor, method or class
weight [arxiv:2605.30099 §IV.B]. Cards: [Blanco-Gonzalez et al. (AI in drug
discovery review)](/papers/2212.08104) and [Teye et al. (emotion detection)](/papers/2605.30099).

The ensemble in [arxiv:1804.02386 §3.2.6] trains seven copies of one network,
each on a bootstrap sample of the training segments drawn with replacement, and
averages their outputs. This is resampling for an ensemble, not a transform of
examples. The card carries this concept tag, but it describes no augmentation
of trajectories, so the paper is listed here for scope only. Card:
[Dabiri & Heaslip (GPS transport modes)](/papers/1804.02386). Its ensemble moves
test accuracy from 82.3% to 84.8% [arxiv:1804.02386 tab. 4]. The table has no
row with one bootstrap copy alone, so that number cannot be credited to
resampling rather than to averaging.

## Papers

- [Ghiasi et al. (Copy-Paste)](/papers/2012.07177): pastes random instance subsets between two jittered training images, with mask and box updates for covered objects, for instance segmentation.
- [Laroca et al. (ALPR)](/papers/1802.09567): colour-inverted and flip-invariant character samples for a plate recogniser, with the valid flips listed per dataset.
- [Liu et al. (graph SSL survey)](/papers/2103.00111): taxonomy of graph pretext tasks, including feature- and topology-level augmentations as contrastive views.
- [Alom et al. (IRRCNN)](/papers/1811.04241): generic rotation, shift, shear, zoom and flip chain for histopathology, with a with and without comparison.
- [Chowdary et al. (face mask)](/papers/2009.08369): eight training-image operations for transfer-learned InceptionV3 on a simulated mask dataset.
- [Yang et al. (distribution calibration)](/papers/2101.06395): training-free feature sampling from borrowed Gaussians for few-shot classification on frozen features.
- [Dai et al. (FSRA)](/papers/2201.09206): repeated augmented satellite copies per step for drone-to-satellite geo-localisation.
- [Park & Kim (ViT analysis)](/papers/2202.06709): analysis of vision transformers, with the strong augmentation recipe used in its experiments.
- [Eftekhar et al. (Omnidata)](/papers/2110.04994): synthetic multi-task dataset rendered from 3D scans; listed for scope, not as an augmentation method.
- [Azimi et al. (steel microstructure)](/papers/1706.06480): rotated 1000 × 1000 px patches with class-balanced strides for pixel-wise steel microstructure segmentation.
- [Dabiri & Heaslip (GPS transport modes)](/papers/1804.02386): seven-model bootstrap ensemble for GPS mode inference; listed for scope, not as an augmentation method.
- [Uddin et al. (SaliencyMix)](/papers/2006.01791): pastes a patch centred on the most salient pixel of one image onto another and mixes labels by patch area, for CNN classification.
- [Thakur et al. (Augmented SBERT)](/papers/2010.08240): recombines gold sentences into new pairs, labels them with a cross-encoder, and trains a bi-encoder on the result.
- [Azizi et al. (synthetic ImageNet)](/papers/2304.08466): fine-tuned text-to-image generator with class-name prompts, used to add synthetic images to ImageNet training.
- [Xiong et al. (TG-LLM)](/papers/2401.06853): four perturbations of synthetic temporal graphs, applied with the same entity and year maps to every field of a training example.
- [Yaseen (YOLOv8 review)](/papers/2408.15857): names mosaic and mixup as YOLOv8 training steps; a review with no experiments of its own.
- [Islam et al. (rice variety)](/papers/2609.10524): flips and rotations that grow a grain photo set sixfold before training ten CNNs.
- [Dai & Adel (NER augmentation)](/papers/2010.11683): four token-level methods for BIO-tagged sequences that keep tags valid.
- [Jiang et al. (token labeling)](/papers/2104.10858): MixToken, a CutMix-style mix on patch embeddings with dense token targets.
- [Sirko et al. (building detection)](/papers/2107.12283): mixup on the segmentation loss only, with separate label maps.
- [Abou Baker et al. (transfer learning)](/papers/2605.11989): offline rotated and noisy copies of a small smartphone photo set.
- [Shanmugam et al. (TTA aggregation)](/papers/2011.11156): learned weights over test-time augmented views; listed for scope, not as training augmentation.
- [Kamilaris & Prenafeta-Boldú (agriculture survey)](/papers/1807.11809): survey of deep-learning studies in agriculture, with the augmentations they used tallied; no experiments of its own.
- [Eldele et al. (TS-TCC)](/papers/2106.14112): weak and strong perturbations of time series as the two views of a contrastive pretraining step.
- [Zhang et al. (NLP technology entities)](/papers/2606.29836): tagger-filtered sentences from unlabelled abstracts for scientific-entity recognition.
- [Shen et al. (mammography)](/papers/1708.09427): label-safe random transforms drawn on the fly for whole screening mammograms.
- [Abou Baker et al. (ewaste)](/papers/2606.23286): reflections, shifts, scales and shears that expand a small smartphone-photo set nine-fold before the split.
- [Min et al. (pre-trained LM survey)](/papers/2111.01243): survey of pre-trained language models, including model-written training data; no experiments of its own.
- [Salloom et al. (water demand)](/papers/2512.06714): one virtual value between consecutive demand readings for GRU forecasting, with k-means class features.
- [Lazanas & Kampouropoulos (multi-generator GAN)](/papers/2607.19153): one WGAN-GP per failure type, with pooled synthetic failures for tabular predictive maintenance.
- [Roth et al. (random view aggregation)](/papers/1505.03046): random cubes around each detection candidate, scored by a ConvNet and averaged; listed for scope, not as training augmentation.
- [Olsen et al. (DeepWeeds)](/papers/1810.05726): ImageNet-pretrained ResNet-50 and Inception-v3 fine-tuned on field weed photos with a heavy transform chain.
- [Somepalli et al. (SAINT)](/papers/2106.01342): row-attention transformer for tables, pre-trained with CutMix and mixup views and a contrastive loss.
- [Li Jing et al. (dimensional collapse)](/papers/2110.09348): shows that contrastive embeddings collapse into a low-dimensional subspace, and ties strong augmentation to it in linear models; proposes DirectCLR.
- [Rakhlin et al. (breast histology)](/papers/1802.00752): 50 colour and 20 crop views per slide, pooled over frozen ImageNet encoders, with a LightGBM classifier.
- [Wang & Qi (CLSA)](/papers/2104.07713): strong views trained to match the similarity profile of their weak twin over a key queue, added to contrastive pretraining.
- [Li, Hou & Che (NLP augmentation survey)](/papers/2110.01852): survey of text augmentation into paraphrasing, noising and sampling; no experiments of its own.
- [dos Santos & Papa (CNN regularization survey)](/papers/2201.03299): survey of CNN regularizers as input, internal and label methods; compiled results not reproduced.
- [Yang et al. (image augmentation survey)](/papers/2204.08610): survey of image augmentation with its own with and without runs on segmentation, classification and detection.
- [Kong et al. (LaserMix++)](/papers/2405.05258): LiDAR band swap between scans with a teacher, extended with camera correspondences and distillation.
- [Safieh et al. (Jordanian speech)](/papers/2609.24410): noisy-student loop for speech recognition, with Gaussian noise, time stretch and pitch shift at each generation.
- [Song et al. (few-shot survey)](/papers/2205.06743): survey that groups few-shot augmentation into hand-crafted and learned methods; no experiments of its own.
- [Blanco-Gonzalez et al. (AI in drug discovery review)](/papers/2212.08104): narrative review that names synthetic data as a remedy for scarce data; no method or experiment.
- [Khan et al. (imbalance benchmark)](/papers/2304.02858): nine resamplers paired with nine ensemble models on imbalanced binary and multi-class sets.
- [Teye et al. (emotion detection)](/papers/2605.30099): names data augmentation and oversampling without specifying them.
- [Hao et al. (raindrop removal)](/papers/2609.37870): ray-traced raindrops with exact masks rendered onto Cityscapes photos, and a three-stage detect, reconstruct and refine network.
- [Howard & Gugger (fastai)](/papers/2002.04688): layered deep learning library whose batch transforms run on the GPU; augmentation is one part of a training API, with no accuracy study.
- [Zhu et al. (GCA)](/papers/2010.14945): graph contrastive learning with edge and feature-dimension drops set by node centrality.
- [Suresh et al. (AD-GCL)](/papers/2106.05819): a GNN augmenter learns per-edge drop probabilities against a GIN encoder for graph contrastive learning, with a penalty on the drop ratio.
- [Wang et al. (video saliency)](/papers/1702.00871): synthetic video made by warping annotated still images with superpixel flow, used to train a fully convolutional saliency network.
- [Vaze et al. (open-set recognition)](/papers/2110.06207): maximum-logit scoring with a strong closed-set recipe that uses RandAugment and label smoothing.
- [Lyu et al. (RTMDet)](/papers/2212.07784): real-time one-stage detector family, with cached Mosaic and MixUp and a large-scale jittering stage.
- [Ji et al. (OXE-AugE)](/papers/2512.13100): cross-paints 16 Open X-Embodiment datasets onto up to nine robot arms by segmentation, inpainting and simulation replay.
- [Yan et al. (InstruGen)](/papers/2411.11394): route-and-instruction pairs from house-tour videos, with instructions written by a multimodal model and checked against the route.
- [Yan et al. (OLiDM)](/papers/2412.17226): object-first LiDAR diffusion, with generated foreground objects placed into a range-image scene model and a per-category loss.
- [Wang & Wang (two-stream skeleton RNN)](/papers/1704.02581): random 3D rotation, scaling and shear of skeleton sequences for a two-stream RNN action recogniser.
- [Zhao et al. (self-consistency deepfake)](/papers/2012.09311): stitched face forgeries with elastic masks, used as training data for a patch-consistency deepfake detector.
- [Zhang et al. (non-Lambertian depth)](/papers/2408.06083): random tone-mapping brightness curves and a gradient-domain loss for depth on glass and mirror surfaces.