---
kind: concept
slug: object-detection
title: Object detection
summary: >-
  Find the objects in an image and give each a class, a box and a score. Variants: two-stage region proposals, single-stage dense detectors, DETR-style set prediction, label assignment for tiny objects, and 3D LiDAR or BEV detection. Latency, hardware, data and training budget decide the choice. The papers use different datasets and metrics, so their numbers cannot be ranked.
tags: [computer-vision, object-detection]
papers: ["1610.06136", "1702.00871", "1705.02544", "1708.00786", "1712.00960", "1712.02294", "1712.06679", "1801.09454", "1802.07856", "1802.09567", "1805.04953", "1807.11809", "1810.08705", "1811.05588", "1903.08701", "1905.12365", "1906.06310", "1907.06781", "1908.08297", "2005.01385", "2011.04841", "2011.07631", "2011.10881", "2012.08548", "2101.06679", "2101.10861", "2102.09480", "2102.10274", "2103.01100", "2103.01415", "2103.02603", "2103.04011", "2103.05950", "2103.07733", "2103.10213", "2103.13137", "2103.14030", "2104.00678", "2104.01111", "2104.10330", "2104.11892", "2105.01922", "2105.03889", "2105.09451", "2105.12931", "2106.08322", "2106.10823", "2106.13797", "2107.02963", "2108.11250", "2109.02497", "2109.07644", "2110.00061", "2110.06922", "2111.08897", "2112.11798", "2202.07008", "2203.17270", "2204.06806", "2204.07974", "2204.08610", "2205.10536", "2205.13790", "2205.14942", "2206.01653", "2206.09474", "2206.15398", "2207.13080", "2208.08738", "2208.11112", "2209.01206", "2209.08538", "2211.12860", "2212.07784", "2303.07601", "2304.00501", "2304.03198", "2304.05170", "2305.11175", "2310.11513", "2311.02877", "2311.06242", "2405.14458", "2405.14747", "2406.19407", "2408.15857", "2408.15865", "2409.19648", "2410.17725", "2410.19277", "2410.19869", "2411.00201", "2411.10291", "2503.01785"]
related: [anchor-free-detection, domain-adaptation, bounding-box-regression-loss, multimodal-fusion, density-map-regression, uncertainty-estimation, gaussian-mixture-model, on-device-inference, robot-simulation, network-pruning, post-training-quantization, depthwise-separable-convolution, label-assignment, object-queries, bipartite-matching, set-prediction, license-plate-recognition, image-manipulation-detection, collaborative-perception, instance-segmentation, knowledge-distillation, semi-supervised-learning, bev-representation, bird-eye-view-perception, oriented-object-detection, vision-transformer, attention-mechanism, graph-neural-network, efficient-vision-backbone, dataset-curation, long-tailed-recognition, deformable-convolution, multi-scale-feature-fusion, focal-loss, visual-object-tracking, benchmark-validity, sim-to-real-transfer, channel-attention, convolutional-neural-network, out-of-distribution-detection, continual-learning, contrastive-learning, data-augmentation, multimodal-language-model, low-rank-adaptation, visual-grounding, salient-object-detection, segmentation-evaluation, multi-task-learning, few-shot-learning, tracking-by-detection, multi-person-pose-estimation, reverse-attention]
status: draft
updated_at: 2026-11-01
---

## Overview

An object detector outputs, for each object it finds, a class label, a box and
a confidence score. Training matches predicted boxes to annotated ones, and
evaluation counts a match when the overlap (IoU) passes a threshold, typically
0.5 [arxiv:1802.09567 §5] [arxiv:1801.09454 §5].

The invariant every variant keeps: the output is a set of labelled boxes, and
each annotated object should be explained by one of them. The variants differ in
how candidate boxes are generated, how each object is matched to candidates
during training, and how duplicates are removed. Dense detectors generate many
candidates and suppress duplicates with non-maximum suppression (NMS).
Set-prediction detectors learn a fixed set of query slots and match them to the
ground truth during training [arxiv:2211.12860 §1] [arxiv:2011.10881 §1].

The cards name SSD, YOLO, Faster R-CNN, RetinaNet, EfficientDet, ATSS, FCOS and
DETR as published detectors, and in 3D they name BEVFormer, DETR3D and
PointPillars. The wiki has no page for those models yet, so they are described
here only through the cards that use them.

## Variants

### Two-stage region proposals (Faster R-CNN)

A region proposal network first proposes candidate boxes from the backbone
features, then a second stage classifies each proposal and regresses its box.
The variant in [Learning Rich Features for Image Manipulation Detection](/papers/1805.04953)
is a Faster R-CNN with two ResNet-101 streams. The proposals come from the RGB
stream only, because proposals from the noise stream performed worse
[arxiv:1805.04953 §3.1] [arxiv:1805.04953 tab. 1]. Test-time NMS uses an IoU
threshold of 0.2 [arxiv:1805.04953 §3.4]. The detector outputs boxes, not
masks, so it targets tampered regions, which need not be objects
[arxiv:1805.04953 §1] [arxiv:1805.04953 §3.4].

- The two streams are fused after RoI pooling, which gives a pixel F1 of 0.722
  on NIST16, against 0.567 for the RGB stream alone and 0.625 for late fusion
  [arxiv:1805.04953 tab. 3].
- Two full ResNet-101 trunks mean backbone compute roughly doubles compared
  with one stream. The paper does not report runtime [arxiv:1805.04953 §3.2].

Card: [RGB-N (Zhou et al.)](/papers/1805.04953).

The same detector family is the base of three other cards, each changing a
different part of training.

- **Tiny-object label assignment.** [RFLA](/papers/2208.08738) replaces the
  overlap test used to pick positive locations with a distance between Gaussian
  receptive fields and Gaussian object models, followed by a second top-1 pass
  [arxiv:2208.08738 §3.1] [arxiv:2208.08738 §3.3]. On AI-TOD, Faster R-CNN
  rises from 11.1 to 21.1 AP, and DetectoRS with RFLA reaches 24.8 AP
  [arxiv:2208.08738 tab. 5]. The setup is ResNet-50, 12 epochs, one RTX 3090
  [arxiv:2208.08738 §4.2].
- **Semi-supervised training.** [Unbiased Teacher](/papers/2102.09480) trains a
  Faster R-CNN with FPN and ResNet-50 on pseudo-labels from an exponential-moving-
  average teacher, with a focal loss on the RoI classifier [arxiv:2102.09480
  §3.2] [arxiv:2102.09480 §3.3]. With 1% of COCO labels it reports 20.75 mAP,
  against 13.97 for STAC [arxiv:2102.09480 tab. 1].
- **Distillation.** [DIST](/papers/2205.10536) is a student-teacher loss for
  dense outputs. On COCO, a Faster R-CNN R50 student goes from 38.4 AP alone to
  40.4 with DIST and 41.8 with DIST plus feature mimicking, against 39.7 with
  classic KD [arxiv:2205.10536 tab. 6]. The same table gives 41.7 for CWD, so DIST
  is not the best of the methods compared there.

### Single-stage dense detectors (SSD, YOLO, RetinaNet)

A single network predicts classes and box offsets on a grid of anchors or
points, with no second per-proposal stage. The cards in this group show the
trade-off in three settings: a small detector for a fixed set of classes, a
cascade of small detectors for a whole pipeline, and a multi-scale fusion
change to SSD.

[Road Damage Detection Using Deep Neural Networks with Images Captured Through a Smartphone](/papers/1801.09454)
trains SSD with MobileNet and SSD with Inception V2 at 300×300 input, on 9,053
dashboard-camera images with eight damage classes [arxiv:1801.09454 §4]
[arxiv:1801.09454 §3.4]. MobileNet runs in 30.6 ms per image on a GPU. The same
model takes 1.5 s on a Nexus 5X phone, measured with an older TensorFlow graph
[arxiv:1801.09454 tab. 3].

[Real-time ALPR with YOLO (Laroca et al.)](/papers/1802.09567)
chains four YOLO-style detectors: vehicle, plate, characters, then letter or
digit. The plate detector is Fast-YOLO [arxiv:1802.09567 §4.1] [arxiv:1802.09567
tab. 1], and character segmentation uses CR-NET, a Fast-YOLO-inspired network
[arxiv:1802.09567 §4.2]. Vehicle detection on
the moving-camera dataset reached at most 97.33% recall with Fast-YOLO, so the
authors switched that stage to YOLOv2 [arxiv:1802.09567 §5.2.1]. The full
pipeline runs at 47 FPS on a Titan Xp for the static-camera dataset
[arxiv:1802.09567 tab. 5] and 35 FPS on the moving-camera dataset
[arxiv:1802.09567 tab. 7].

[FSSD](/papers/1712.00960) keeps SSD's structure and fuses three VGG16 maps
(conv4_3, fc7, conv7_2) into one feature map before the detection pyramid
[arxiv:1712.00960 §3.1] [arxiv:1712.00960 §4.2]. The authors report 78.8 mAP on
VOC2007 test at 65.8 FPS on a 1080Ti, and 82.7 with COCO pre-training
[arxiv:1712.00960 tab. 3]. They say the fusion costs a small amount of speed,
without a number in the card [arxiv:1712.00960 §4].

- Per class, recall is uneven. In the road-damage baseline the rarest classes,
  D11 and D40, have low recall, which the authors link to few training examples
  [arxiv:1801.09454 §5] [arxiv:1801.09454 tab. 2].
- A weak early stage caps the whole pipeline. The only vehicle missed on the
  moving-camera dataset was lost in the plate stage, where a false positive
  outscored the real plate [arxiv:1802.09567 §5.2.2].
- Speed and accuracy are both reported on one colonoscopy benchmark
  [arxiv:2011.07631 tab. III]. YOLOv4 reaches AP 0.8513 at 48 FPS and RetinaNet
  with ResNet-101 reaches 0.8745 at 16.8 FPS. Faster R-CNN with ResNet-50 reaches
  0.7866 at 8 FPS, and EfficientDet-D0 reaches 0.4756 at 35 FPS. The authors name
  YOLOv4 as the best detector for the balance of accuracy and speed
  [arxiv:2011.07631 §VI].

[YOLO-LITE](/papers/1811.05588) is a seven-layer network with no batch
normalisation, found by trial and error from Tiny-YOLOv2 for computers without a
GPU [arxiv:1811.05588 §3.1] [arxiv:1811.05588 tab. VII]. On PASCAL VOC 2007
validation it reaches 33.57% mAP at 21 FPS on one laptop without a GPU, against
40.48% and 2.4 FPS for Tiny-YOLOv2 [arxiv:1811.05588 tab. III]. Its COCO model
reaches 12.26% mAP at the same speed [arxiv:1811.05588 tab. V]. Speed is
measured on one device, and the authors call COCO-level accuracy the limit for
real applications [arxiv:1811.05588 §7].

[microYOLO](/papers/2408.15865) is a YOLO-style detector for a Cortex-M7
microcontroller with no accelerator. It takes 128 x 128 input, uses a small
depthwise-separable backbone, and is then pruned with an L1 criterion and
quantized to 8 bits [arxiv:2408.15865 §2]. On the OpenMV H7 R2 the fridge model
uses 752 KB of Flash and 324 KB of RAM at 3.46 FPS [arxiv:2408.15865 tab. 2].
Author-reported mAP50 is 56.4% on the fridge task, 27.7% on humans and 12.3% on
vehicles [arxiv:2408.15865 §3.1] [arxiv:2408.15865 fig. 1]. Shrinking the vehicle
input to 88 x 88 cuts Flash to 133 KB and RAM to 152 KB, and raises speed to
8.40 FPS [arxiv:2408.15865 tab. 3]. The authors call the work in progress
[arxiv:2408.15865 §1].

[Edge YOLO](/papers/2205.14942) is a slimmed YOLOv4-style detector for a Jetson
board in a vehicle, with a cloud server that retrains it and sends new weights
back [arxiv:2205.14942 §III-A]. Its CSPDarknet53 backbone is cut to 17 layers
from 88, the neck is an FPN without PANet, and there are three heads at 13, 26
and 52 cells [arxiv:2205.14942 §III-B] [arxiv:2205.14942 tab. 1]. On COCO2017 it
reports 0.473 mAP@0.5 at 26.6 FPS on Xavier, against 0.657 at 2.1 FPS for
YOLOv4 [arxiv:2205.14942 tab. 5]. Under TensorRT FP16 the same table gives 0.325
at 49.5 FPS. The 26.6 FPS figure comes from a run whose precision the text does
not state [arxiv:2205.14942 tab. 5]. The Soft-NMS threshold, σ and decay
function are not stated, so the post-processing cannot be reproduced exactly
[arxiv:2205.14942 alg. 2].

[DDOD](/papers/2107.02963) keeps a one-stage FPN detector and separates three
parts of its training: classification and regression choose their own positives
with a cost-based assignment, the first convolution of each head branch is
deformable, and each pyramid level is weighted by how many positives it has
received [arxiv:2107.02963 §4.1] [arxiv:2107.02963 §4.2] [arxiv:2107.02963 §4.3].
On ATSS, the three parts together raise COCO val AP from 39.4 to 41.6, and each
part alone adds 0.6 to 1.2 AP [arxiv:2107.02963 tab. 4]. The same recipe gives
+2.0 to +2.4 AP on RetinaNet, FCOS and ATSS [arxiv:2107.02963 tab. 3]. The
released code runs the assignment twice per iteration, and the paper gives no
matched speed comparison [arxiv:2107.02963 tab. 8].

[YOLO-Z](/papers/2112.11798) changes YOLOv5 for small objects, such as traffic
cones seen from a racing car. Its strongest change replaces the lowest-resolution
feature map with a higher-resolution one, which the authors call XS_ex
[arxiv:2112.11798 §3.3.3]. At the S scale, small-object mAP@0.5 rises from 0.869
to 0.925 while inference goes from 8 to 8.9 ms on a GTX 1080 Ti [arxiv:2112.11798
tab. 2]. Feature-map changes gave the largest small-object gain at every scale
[arxiv:2112.11798 §4.3]. The results come from one dataset, with single runs.

[YOLO5Face](/papers/2105.12931) treats face detection as generic object detection.
It adds a five-point landmark head to YOLOv5, a stride-64 output for large faces,
and ShuffleNetV2 backbones for the smallest models [arxiv:2105.12931 §3.2]. On
WiderFace validation, the largest model reports 96.67, 95.08 and 86.55 (Easy,
Medium, Hard), and the 1.7M-parameter YOLOv5n reports 93.61, 91.54 and 80.53
[arxiv:2105.12931 tab. 3]. The paper gives FLOPs and parameters, not measured
latency [arxiv:2105.12931 §5].

Cards: [Road Damage Detection](/papers/1801.09454),
[Automatic License Plate Recognition](/papers/1802.09567),
[FSSD](/papers/1712.00960),
[Real-Time Polyp Detection, Localization and Segmentation](/papers/2011.07631),
[YOLO-LITE](/papers/1811.05588), [microYOLO](/papers/2408.15865), [DDOD](/papers/2107.02963),
[YOLO-Z](/papers/2112.11798), [YOLO5Face](/papers/2105.12931), [Edge YOLO](/papers/2205.14942).

### Box regression loss on auxiliary boxes (Inner-IoU)

The box regression loss sits on the box head of any dense or two-stage detector,
so this change does not depend on the architecture. [Inner-IoU](/papers/2311.02877)
replaces only the IoU term of a CIoU or SIoU loss. It scales the predicted and
ground-truth boxes about their own centres, takes the IoU of the scaled boxes,
and keeps the penalty terms on the full boxes [arxiv:2311.02877 §3.2]
[arxiv:2311.02877 eq. 15-27]. The authors argue that a ratio below 1 sharpens the
gradient for high-IoU samples and a ratio above 1 does the same for low-IoU
samples [arxiv:2311.02877 §3.1].

- On VOC with YOLOv7-tiny, Inner-CIoU at ratio 0.70 reaches 38.38 mAP50:95
  against 37.64 for CIoU, and Inner-SIoU at 0.75 reaches 38.52 against 37.31 for
  SIoU [arxiv:2311.02877 tab. 1]. Every VOC row is above its baseline, by 0.61 to
  1.21 points of mAP50:95 [arxiv:2311.02877 tab. 1].
- On AI-TOD with YOLOv5s the gains are mixed. At ratio 1.10 mAP50:95 falls by
  0.17 while AP50 rises by 0.72; at ratios 1.13 and 1.15 mAP50:95 rises by 0.17
  [arxiv:2311.02877 tab. 2].
- The paper gives no rule for choosing the ratio and no validation split, and each
  row is a single run [arxiv:2311.02877 §4.2]. Only CIoU and SIoU were tested,
  although GIoU, DIoU and EIoU are written into the method [arxiv:2311.02877
  §4.2] [arxiv:2311.02877 eq. 23-27].
- With a ratio below 1, two overlapping boxes can have disjoint auxiliary boxes.
  The IoU term then has zero gradient, a case the paper does not discuss
  [arxiv:2311.02877 eq. 19] [code: innerciou.py#L16-L17].

Card: [Inner-IoU](/papers/2311.02877).

### Real-time one-stage design study (RTMDet)

[RTMDet](/papers/2212.07784) is a family of one-stage detectors, from tiny to
extra-large, built by testing design choices one at a time for real-time use
[arxiv:2212.07784 §1]. Three choices carry most of the design. The basic block
uses a 5 by 5 depthwise convolution, which the authors prefer over 3 by 3 and
7 by 7 for its speed-accuracy balance [arxiv:2212.07784 tab. 5]. The network
has fewer blocks per stage and wider channels, and the neck gets a larger
expansion ratio so that backbone and neck have similar capacity
[arxiv:2212.07784 §3.2]. Training uses a dynamic label assignment in which the
classification target is the IoU between prediction and ground truth, and the
regression cost is the negative log of that IoU [arxiv:2212.07784 §3.3]
[arxiv:2212.07784 eq. 2] [arxiv:2212.07784 eq. 3].

- Reported on COCO val2017 with the authors' own latency setup: FP16,
  TensorRT, batch 1, one NVIDIA 3090, NMS excluded [arxiv:2212.07784 §4.1].
  RTMDet-x reaches 52.8 AP at 3.10 ms. RTMDet-s reaches 44.6 AP at 1.22 ms,
  against 43.5 AP at 0.92 ms for YOLOv6-s, which the table lists with a
  letterboxed input [arxiv:2212.07784 tab. 2]. So RTMDet-s is more accurate and
  slower than that YOLOv6-s row, not faster.
- Mosaic and MixUp need several images per sample, which costs loading time.
  Keeping a cache of already-loaded images cuts that cost, with cached Mosaic
  at 24.0 ms per 100 images against 87.1 ms uncached [arxiv:2212.07784 §3.3]
  [arxiv:2212.07784 tab. 7]. A large cache with random eviction behaves like
  uncached augmentation, while a small cache with first-in-first-out eviction
  repeats images within a batch [arxiv:2212.07784 §3.3].
- The same family extends to instance segmentation and to rotated boxes, each
  with small changes to the output part of the network [arxiv:2212.07784 §3.4].
  The paper's rotated results are on DOTA [arxiv:2212.07784 tab. 4].

### NMS-free YOLO (YOLOv10)

[YOLOv10](/papers/2405.14458) keeps the YOLO layout but removes NMS at
inference. It trains two heads with the same structure: a one-to-many head with
rich supervision, and a one-to-one head that makes the final predictions. A
shared matching metric keeps the two consistent [arxiv:2405.14458 §3.1]. The
authors then cut cost in several components, stage by stage [arxiv:2405.14458
§3.2]. On COCO val, the N to X models reach 38.5 to 54.4 AP, with end-to-end
latency from 1.84 to 10.70 ms on a T4 with TensorRT FP16 [arxiv:2405.14458 tab.
1]. The paper's own remaining gap is 1.0 AP for N and 0.5 AP for S, against
one-to-many training [arxiv:2405.14458 §4.3] [arxiv:2405.14458 app. A.9].

### YOLO family comparisons and reviews

Six cards compare or review YOLO versions. They do not share a protocol, so
their rankings are not directly comparable.

- [YOLO Evolution](/papers/2411.00201) benchmarks 33 Ultralytics checkpoints, from
  YOLOv3u to YOLOv12, on three datasets: traffic signs, African wildlife and
  ships. All use one training setting, with one run per model [arxiv:2411.00201
  §3.2] [arxiv:2411.00201 tab. 3]. Its traffic-sign table puts YOLO11m at mAP50
  0.893 and mAP50-95 0.795 [arxiv:2411.00201 fig. 12]. Averaged over scales and
  datasets, YOLO11 and YOLOv12 tie for first on accuracy, and YOLOv12 ranks last
  on speed [arxiv:2411.00201 tab. 4].
- [YOLO11 and YOLOv8 on orchard fruit](/papers/2410.19869) trains YOLO11-seg and
  YOLOv8-seg in five sizes each, on 991 images of immature apples [arxiv:2410.19869
  §3.2]. YOLO11m-seg has the best box mAP@50 overall, 0.876, and YOLOv8n-seg the
  lowest inference time, 3.3 ms [arxiv:2410.19869 tab. 2] [arxiv:2410.19869 tab.
  3]. Post-processing differs widely: 56.7 ms for YOLO11n-seg against 2.6 ms for
  YOLO11s-seg. The speed ranking therefore depends on which stage is measured.
- [What is YOLOv8](/papers/2408.15857) is a descriptive review. It runs no
  experiments, and its YOLOv5 against YOLOv8 table gives no dataset or hardware
  [arxiv:2408.15857 tab. 1]. Use it for architecture, not for numbers.
- [A review of YOLO architectures from YOLOv1 to YOLOv8](/papers/2304.00501)
  covers sixteen versions, with a summary table of reported COCO AP. The table
  gives 33.0 for YOLOv3 and 55.8 for YOLOv5, while the text gives 36.2 for
  YOLOv3-spp and 50.7 for YOLOv5 at 640 pixels [arxiv:2304.00501 tab. 4]
  [arxiv:2304.00501 §6.3] [arxiv:2304.00501 §9]. The paper does not run any model
  itself [arxiv:2304.00501 §8].
- [YOLO11 architectural overview](/papers/2410.17725) describes C3k2 and C2PSA
  blocks and one model family from nano to xlarge. It runs no experiments. Its
  accuracy and latency figures are read from a third-party chart and given as
  approximations [arxiv:2410.17725 §6] [arxiv:2410.17725 fig. 2].
- [YOLO advances to its genesis](/papers/2406.19407) reviews YOLOv1 to YOLOv12
  and five variant detectors, walking the versions in reverse order. It runs no
  experiments, so every number is from a cited work [arxiv:2406.19407 abstract].
  The paper states that no datasets were generated or analysed [arxiv:2406.19407
  p. 68]. Its keywords name YOLOv13 to YOLOv20, which the body does not cover, so
  those versions are not reviewed here [arxiv:2406.19407 abstract].

### Set prediction with transformers (DETR family)

Each object is assigned to one query, and the model learns to output one box per
query with no NMS at inference. The cards disagree on why DETR-style training is
slow, and they propose different fixes.

[TSP-FCOS and TSP-RCNN](/papers/2011.10881) remove DETR's decoder and run a
six-layer transformer encoder over about 700 features of interest, which come
from FCOS or an RPN [arxiv:2011.10881 §4]. The authors name the cross-attention
module as the main cause of slow convergence [arxiv:2011.10881 §1]
[arxiv:2011.10881 §3.2]. With ResNet-50 at 36 epochs they report 43.1 and 43.8
COCO val AP, against 42.0 for DETR after 500 epochs [arxiv:2011.10881 tab. 1].

[H-DETR](/papers/2207.13080) keeps one-to-one matching for the final model and
adds a training-only group in which each ground truth appears K times, with K=6
for COCO detection. Masked self-attention keeps the two groups apart
[arxiv:2207.13080 §3]. The authors note that Deformable-DETR selects fewer than
30 of its 300 queries as positives per image [arxiv:2207.13080 §1]. Deformable-
DETR with ResNet-50 at 12 epochs goes from 47.0 to 48.7 AP [arxiv:2207.13080
tab. 1]. Inference cost is unchanged [arxiv:2207.13080 §3].

[Co-DETR](/papers/2211.12860) adds auxiliary detection heads (ATSS and Faster R-CNN)
on the encoder output, with one-to-many label assignment, and feeds their
positive boxes back into the decoder as extra query groups
[arxiv:2211.12860 §3.2] [arxiv:2211.12860 §3.3]. Its inference path is the
plain set predictor.

- On Deformable-DETR with a ResNet-50 backbone at 12 epochs, COCO val AP goes
  from 37.1 without auxiliary heads to 42.3 with one ATSS head and 42.9 with two
  heads [arxiv:2211.12860 tab. 2].
- The extra training cost is substantial. For Deformable-DETR++ with ResNet-50,
  GPU hours go from 70 to 86 with one head, 120 with two, and 280 with six
  [arxiv:2211.12860 tab. 7].
- The strongest published Co-DINO results use NMS (soft-NMS in the released
  configs), so they are not NMS-free [arxiv:2211.12860 §4.3].

[OrientedFormer](/papers/2409.19648) applies a DETR-style decoder to oriented
boxes in remote-sensing images. It encodes each box as a Gaussian for the
positional encoding, biases decoder self-attention with Gaussian Wasserstein
distances, and samples cross-attention features at points rotated by the box
angle [arxiv:2409.19648 §3.3] [arxiv:2409.19648 §3.4] [arxiv:2409.19648 §3.5].
With ResNet-50 it reports 67.28 AP50 on DIOR-R and 75.37 on DOTA-v1.0
single-scale, against 66.12 and 74.16 for ARS-DETR [arxiv:2409.19648 tab. 3]
[arxiv:2409.19648 tab. 4]. It trains for 12 epochs against 36 for two
end-to-end baselines, but uses about 2.8 times their FLOPs [arxiv:2409.19648
§4.8] [arxiv:2409.19648 tab. 14].

[PubTables-1M](/papers/2110.00061) applies DETR without change to table images.
One DETR finds the table on the page, with 15 queries, and a second predicts six
structure classes, including rows, columns and spanning cells, with 125 queries
[arxiv:2110.00061 §4] [arxiv:2110.00061 app. Training Details]. On the canonical
test split, it reaches 0.8138 content accuracy against 0.1039 for Faster R-CNN,
whose training details the card records as "default" only [arxiv:2110.00061
tab. 4]. The authors first merge oversegmented header cells, because the same
table otherwise has two valid readings [arxiv:2110.00061 §3]. Scores on canonical
and non-canonical test data are not comparable [arxiv:2110.00061 §5].

Cards: [TSP](/papers/2011.10881), [H-DETR](/papers/2207.13080),
[DETRs with Collaborative Hybrid Assignments Training](/papers/2211.12860),
[OrientedFormer](/papers/2409.19648), [PubTables-1M](/papers/2110.00061).

### Attention-based heads and backbones

These cards keep a dense or two-stage detector and replace one of its parts with
attention. [Dynamic Head](/papers/2106.08322) stacks blocks over a feature
pyramid, viewed as levels, positions and channels. Each block applies three
attentions in turn: one over scales, one over spatial positions through
deformable sampling, and one per channel [arxiv:2106.08322 §3.2]. Plugged into
ATSS with ResNet-50 at 1x, it moves COCO val AP from 39.4 to 42.6 [arxiv:2106.08322
tab. 3]. The gain also appears on Faster R-CNN, RetinaNet, FCOS and RepPoints, at
1.2 to 3.2 AP [arxiv:2106.08322 §4.2] [arxiv:2106.08322 tab. 3]. Six blocks add
21.5 GFLOPs to a 254.39 GFLOP baseline [arxiv:2106.08322 tab. 2].

[Swin Transformer](/papers/2103.14030) is a backbone rather than a detector. It
limits self-attention to local windows and shifts the window grid in every other
block, so neighbouring windows exchange information [arxiv:2103.14030 §3]. Used
with Cascade Mask R-CNN, Swin-T reaches 50.5 box AP on COCO val, against 46.3
for ResNet-50 [arxiv:2103.14030 tab. 2]. The paper's shifted-window ablation
gives 47.7 without shifting and 50.5 with it [arxiv:2103.14030 tab. 4]. Its
speed row is on one V100 in images per second, 755 for shifted windows against
183 for a naive sliding window [arxiv:2103.14030 tab. 5].

[Conformer](/papers/2105.03889) is a hybrid backbone: a CNN branch and a
transformer branch run side by side and exchange features in every block through
a Feature Coupling Unit [arxiv:2105.03889 §3.2]. In Faster R-CNN on COCO val,
Conformer-S/32 reaches 43.1 box AP against 39.4 for ResNet-101 FPN, and in Mask
R-CNN it reaches 39.7 mask AP against 36.1 [arxiv:2105.03889 §4.3]
[arxiv:2105.03889 tab. 3]. The ResNet-101 rows are taken from mmdetection, not
re-run by the authors [arxiv:2105.03889 tab. 3]. Throughput is lower: 13.5 FPS
against 15.9 for ResNet-101 in the Faster R-CNN setting, on a V100 at batch 1
[arxiv:2105.03889 app. C].

[PVT v2](/papers/2106.13797) is a four-stage pyramid transformer backbone. Its
attention pools keys and values to a fixed 7 by 7 grid, so the cost grows
linearly with the number of tokens, while queries still come from every token
[arxiv:2106.13797 §3.2] [arxiv:2106.13797 eq. 2]. With ATSS, the B2 model
reaches 49.9 box AP on COCO val against 47.2 for Swin-T, and the linear B2-Li
reaches 48.9 at 194 GFLOPs against 215 for Swin-T [arxiv:2106.13797 tab. 4].
With RetinaNet at 1x, B4 reaches 46.1 AP against 42.6 for PVT v1-Large
[arxiv:2106.13797 tab. 3]. The comparison table does not state its schedule
[arxiv:2106.13797 tab. 4].

[RFAConv](/papers/2304.03198) replaces a k×k convolution with one whose weights
differ at each window position. Each window gets its own softmax over its k×k
slots, computed from pooled features, so the weights are not shared between
overlapping windows [arxiv:2304.03198 §3.3]. In YOLO models the authors report
COCO2017 AP rising from 36.4 to 37.7 for YOLOv8n and from 27.5 to 29.0 for
YOLOv5n [arxiv:2304.03198 tab. 7]. Per-image validation time for YOLOv8n goes
from 4.2 to 4.5 ms [arxiv:2304.03198 tab. 7]. The expanded tensor is k² times the
input, a memory cost the authors state [arxiv:2304.03198 §4.6].

Cards: [Dynamic Head](/papers/2106.08322), [Swin Transformer](/papers/2103.14030),
[Conformer](/papers/2105.03889), [PVT v2](/papers/2106.13797),
[RFAConv](/papers/2304.03198).

### Language-model decoders (VisionLLM)

[VisionLLM](/papers/2305.11175) writes detection as text for a language model
to generate. A language-guided tokenizer turns the image into 100 learned
tokens, shaped by the instruction. An Alpaca-7B decoder, adapted with LoRA,
then writes class and coordinate tokens in the format the instruction names
[arxiv:2305.11175 §3.3] [arxiv:2305.11175 §3.4] [arxiv:2305.11175 §4.1]. A box
is two points, the top-left and bottom-right corners, and the category set is
given in the instruction [arxiv:2305.11175 §3.2] [arxiv:2305.11175 fig. 9].

- With a ResNet-50 backbone the model reaches 44.6 box AP on COCO, 1.4 above
  Pix2Seq at the same backbone. With InternImage-H it reaches 60.2
  [arxiv:2305.11175 tab. 1]. The ResNet-50 model trained separately for
  detection reaches 44.8 [arxiv:2305.11175 tab. 1].
- Masks are written as boundary points, and mask AP rises with their number:
  18.5 with 8 points against 25.1 with 24 [arxiv:2305.11175 tab. 2]. Box
  detection uses two points, so this trade-off does not apply to boxes.
- The authors argue that parallel decoding is more efficient than token-by-token
  output, but report no latency or throughput numbers [arxiv:2305.11175 §4.2].
  The card records that the linked repository has no implementation of this
  method at the pinned commit, so the numbers cannot be checked against code
  [card](/papers/2305.11175).

### Oriented boxes in aerial images

[ReDet](/papers/2103.07733) makes an oriented-box detector rotation-equivariant
instead of relying on rotated copies of the training data. Its backbone is a
ResNet-50 with FPN, equivariant to eight discrete rotations, and RiRoI Align pools
rotated proposals after aligning them in space and orientation [arxiv:2103.07733
§5.4] [arxiv:2103.07733 tab. 1]. On DOTA-v1.0 with multi-scale training and
rotation augmentation it reports 80.10 mAP, against 79.42 for the best other row
in its table [arxiv:2103.07733 tab. 6]. Its backbone is 12 Mb against 103 Mb for
the baseline [arxiv:2103.07733 tab. 1]. The invariance is exact only at the grid
angles, because RiRoI Align interpolates between neighbouring orientations
[arxiv:2103.07733 §4.1] [arxiv:2103.07733 §5.3]. For a DETR-style alternative on
the same kind of data, see OrientedFormer above.

Card: [ReDet](/papers/2103.07733).

### 3D detection from point clouds, cameras and LiDAR-camera fusion

Several cards detect objects in 3D. They differ on what the detector sees,
whether it is shared between vehicles, and how many cameras it uses.

[Group-Free 3D](/papers/2104.00678) detects indoor objects from point clouds
without a hand-made grouping step. Each candidate attends to all points through
stacked decoder layers, and the boxes from all layers are pooled and filtered
with NMS at IoU 0.25 [arxiv:2104.00678 §3.2]. The authors report 67.3 mAP@0.25
and 48.9 mAP@0.5 on ScanNet V2 for the best run, and 66.3 and 48.5 averaged over
25 trials [arxiv:2104.00678 tab. 1]. A 2D DETR-style setup on the same data
reaches 39.6 and 21.4 after 400 epochs [arxiv:2104.00678 tab. 10].

[BADet](/papers/2104.10330) is a LiDAR detector that refines its proposals
jointly rather than one at a time. Proposals within a cut-off distance form a
graph, and message passing between them updates each proposal's features. On
KITTI test, Car, its moderate AP3D is 81.61, level with Voxel R-CNN at 81.62
[arxiv:2104.10330 tab. 1]. Its average runtime is 140 ms on a GTX 1080 Ti
[arxiv:2104.10330 §4.5], which the authors give as 7.1 FPS against 1.5 for
Point-GNN [arxiv:2104.10330 tab. 9]. With the graph refiner, KITTI val moderate
AP3D rises from 79.94 to 85.77 over the same model without it [arxiv:2104.10330
tab. 5].

[LaserNet](/papers/1903.08701) runs one fully convolutional network on the LiDAR
range image, with one row per laser, rather than on a bird's-eye-view grid
[arxiv:1903.08701 §3.1]. Each point predicts a mixture of box distributions.
Mean-shift clustering merges the predictions, and an adaptive NMS prunes them
using the predicted spread [arxiv:1903.08701 §3.4] [arxiv:1903.08701 §3.6]. On
the ATG4D validation set, vehicle AP at IoU 0.7 over 0 to 70 m is 85.34, against
80.99 for PIXOR [arxiv:1903.08701 tab. 1]. On KITTI the table gives 73.77 moderate,
against 77.05 for PIXOR, and the authors say their method is worse than current
BEV detectors there [arxiv:1903.08701 tab. 4] [arxiv:1903.08701 §4.4]. Total
runtime is 30 ms on a 1080 Ti, with pre- and post-processing included
[arxiv:1903.08701 tab. 3] [arxiv:1903.08701 §4.3]. The paper does not say how
to obtain ATG4D, so its headline result may not be reproducible from the paper
alone [arxiv:1903.08701 §4.1].

[AVOD](/papers/1712.02294) fuses a LiDAR bird's-eye-view map with an RGB image in
full-resolution feature pyramids. A region proposal network proposes 3D boxes
from both views, and a second stage refines each box and predicts its
orientation [arxiv:1712.02294 §3.3] [arxiv:1712.02294 §3.4]. On KITTI test, Car,
the pyramid version reports 81.94, 71.88 and 66.38 AP3D for easy, moderate and
hard, at about 0.1 s per frame on a TITAN Xp [arxiv:1712.02294 tab. 2]
[arxiv:1712.02294 §4]. The feature-pyramid variant has the highest car scores,
with the base network listed beside it [arxiv:1712.02294 tab. 2]. All numbers
are author-reported.

[BEVFusion](/papers/2205.13790) gives the camera its own stream: image features
are lifted into a bird's-eye-view grid and fused with a LiDAR BEV feature, so
the camera path does not need LiDAR points [arxiv:2205.13790 §3] [arxiv:2205.13790
§3.3]. On nuScenes validation with a TransFusion-L LiDAR stream, fusion gives
67.9 mAP against 64.9 for LiDAR alone [arxiv:2205.13790 tab. 1]. Under simulated
LiDAR loss, the abstract reports margins of 15.7 to 28.9 mAP points over earlier
methods [arxiv:2205.13790 §4.4.1].

[OPV2V](/papers/2109.07644) is a benchmark for cars that share features over a
radio link. Nearby cars share LiDAR features, so a car sees objects hidden
behind other vehicles [arxiv:2109.07644 §3.1]. The benchmark compares four
LiDAR detectors under no fusion, late fusion, early fusion and intermediate
fusion [arxiv:2109.07644 §5.1] [arxiv:2109.07644 tab. 4].

- For VoxelNet on the default towns, AP at IoU 0.7 rises from 0.526 with no
  fusion to 0.864 with intermediate fusion [arxiv:2109.07644 tab. 4].
- Compressing the shared features to 4096× costs about 3% AP and still beats
  early and late fusion. The authors estimate about 5 ms to send one message at
  27 Mbps and 300 m [arxiv:2109.07644 §5.6].
- All results are simulated. The paper does not give a license for the dataset,
  and the released code is non-commercial [arxiv:2109.07644 §1].

[V2V4Real](/papers/2303.07601) is a real-world counterpart to simulated car-to-car
data. Two instrumented cars drove together in Columbus, Ohio, and each car's LiDAR
was labelled with 3D boxes [arxiv:2303.07601 §3.1] [arxiv:2303.07601 §3.2]. Its
detection benchmark uses LiDAR only. There, CoBEVT reaches 66.5 AP@0.5 in the sync
setting, against 39.8 for no fusion [arxiv:2303.07601 tab. 3]. Adapting a model
trained on simulated OPV2V data lifts CoBEVT from 32.6 to 40.2 AP@0.5
[arxiv:2303.07601 tab. 5]. The dataset is not openly licensed, so check its card
before reuse.

[Interpretable motion planner](/papers/2101.06679) feeds LiDAR sweeps and an HD
map through one network that also predicts vehicle detections and motion
[arxiv:2101.06679 §3.1]. Detection is one head of a planner, so its own accuracy
is not the headline; the paper reports a lower future collision rate than the
compared planners on a private dataset [arxiv:2101.06679 §4].

[VoTr](/papers/2109.02497) swaps the sparse-convolution backbone of a voxel
detector for Transformer blocks. Each occupied voxel attends only to other
occupied voxels, at most 48 per query, chosen from a near range and a dilated
range with larger steps [arxiv:2109.02497 §3.2] [arxiv:2109.02497 §3.3]
[arxiv:2109.02497 §4.1]. A GPU hash table finds those neighbours
[arxiv:2109.02497 §3.4]. Its single-stage form, VoTr-SSD, reports 78.25 AP3D
moderate on KITTI test, against 75.96 for SECOND [arxiv:2109.02497 tab. 2]. The
authors' runtime is 14.65 Hz against 20.73 Hz for SECOND [arxiv:2109.02497
tab. 8]. The code repository has no license file, so check the card before
reusing any of it.

Two cards detect objects from several cameras or from cameras and LiDAR together.
[PolarFormer](/papers/2206.15398) predicts objects on a polar grid of distance
and azimuth, fed by six surround cameras [arxiv:2206.15398 §3.1] [arxiv:2206.15398
§3.4]. On nuScenes validation with a ResNet-101 backbone, it reports 39.6 mAP and
45.8 NDS, against 37.5 and 44.8 for BEVFormer-S [arxiv:2206.15398 tab. 3]. Its
50.0 mAP setting uses a VoVNet backbone and extra pretraining data [arxiv:2206.15398
§4]. The paper reports no runtime [arxiv:2206.15398 §4].

[DeepInteraction](/papers/2208.11112) keeps a LiDAR map and the camera features as
separate representations, and lets them exchange information through
projections. A two-layer encoder does the exchange, and a decoder reads the image
and LiDAR maps in turn [arxiv:2208.11112 §3.1] [arxiv:2208.11112 §3.2]. The base
model, with a ResNet-50 image backbone, reports 70.8 mAP and 73.4 NDS on nuScenes
test. Its LiDAR-only baseline, TransFusion-L, reports 65.5 mAP on the same split
[arxiv:2208.11112 tab. 1]. DeepInteraction runs at 2.6 FPS on a V100, against 3.8
for TransFusion [arxiv:2208.11112 tab. 2]. The method depends on sensor
calibration [arxiv:2208.11112 app. A.3].

[CenterFusion](/papers/2011.04841) adds radar to a camera-only, center-point 3D
detector. Radar points are expanded to pillars, the closest one inside each
object's frustum is kept, and its depth and velocity are painted onto image-shaped
maps for second-stage heads [arxiv:2011.04841 §4.2] [arxiv:2011.04841 §4.3]. On
nuScenes validation, both models with DLA, NDS rises from 0.328 to 0.453, while mAP
rises only from 0.306 to 0.332 [arxiv:2011.04841 tab. 1]. On the test split it
reports NDS 0.449 against 0.400 for the Hourglass CenterNet, but mAP 0.326 against
0.338 [arxiv:2011.04841 tab. 1]. The authors' ablation attributes most of the
velocity error reduction to the frustum association [arxiv:2011.04841 tab. 3].
The paper and the released code differ in several parameters, such as the
test-time region enlargement and the pillar size, so read the card before
reproducing numbers [arxiv:2011.04841 §5].

Card: [CenterFusion](/papers/2011.04841).

Two more cards detect from camera images alone, with no LiDAR in the loop, and
both lift image features into a bird's-eye-view layout. [BEVFormer](/papers/2203.17270)
turns six surround-camera images into a BEV grid of 200 by 200 learnable queries
with a six-layer encoder [arxiv:2203.17270 §3.1]. Spatial cross-attention reads
image features where each grid cell's 3D points project, and temporal
self-attention reads the previous BEV map [arxiv:2203.17270 §3.3]
[arxiv:2203.17270 §3.4]. On nuScenes val with ResNet-101 it reports 51.7 NDS
against 44.8 for BEVFormer-S [arxiv:2203.17270 tab. 2]. Its 56.9 NDS on test uses
a VoVNet-99 backbone pre-trained on extra depth data [arxiv:2203.17270 tab. 1].
[DETR3D](/papers/2110.06922) uses no depth network and no NMS. Its 900 learned
queries each hold a 3D reference point, which is projected into every camera to
sample features [arxiv:2110.06922 §3.3]. It reports 0.374 NDS and 0.303 mAP on
nuScenes val with ResNet-101 [arxiv:2110.06922 tab. 1]. Its mATE is worse than
FCOS3D's, and the authors suspect that FCOS3D's direct depth prediction
supervises translation more strongly [arxiv:2110.06922 §4.2]. In a pseudo-LiDAR
baseline the authors built, NDS is 0.160 [arxiv:2110.06922 tab. 4].

[CaDDN](/papers/2103.01100) detects cars from one camera. For every pixel it
predicts a distribution over depth bins rather than a single depth value. Image
features are weighted by those distributions, lifted into a frustum, sampled
into a voxel grid, collapsed to bird's-eye view, and passed to a PointPillars
style head, all trained end to end [arxiv:2103.01100 §3.1] [arxiv:2103.01100
§3.2]. LiDAR gives the depth supervision during training only [arxiv:2103.01100
§3.4]. On KITTI test, Car, moderate AP R40, it reports 13.41 against 9.71 for
M3D-RPN [arxiv:2103.01100 tab. 1]. On KITTI val, adding the depth-distribution
supervision raises moderate Car AP from 6.43 to 14.03 [arxiv:2103.01100 tab. 3],
and training the depth head jointly with the detector gives 13.71, against 12.26
when the depth is estimated separately [arxiv:2103.01100 tab. 4]. The released
code implements the focal weighting differently from the wording in the paper,
so read the card before porting it.

[Pseudo-LiDAR++](/papers/1906.06310) estimates depth from a stereo pair with a
network trained on depth rather than disparity. A few LiDAR points then correct the
depth map, and their depths spread to nearby points through a graph
[arxiv:1906.06310 §3] [arxiv:1906.06310 §4]. The paper argues that a disparity
loss spends effort on near objects, because a fixed disparity error costs far more
depth at long range [arxiv:1906.06310 §3] [arxiv:1906.06310 eq. 5]. On KITTI
validation, car, moderate, IoU 0.7, PointRCNN, AP_BEV rises from 56.0 with the
pseudo-LiDAR baseline to 64.0 with the new stereo network, and to 76.9 with four
simulated LiDAR beams added [arxiv:1906.06310 tab. 1]. On the test split the
four-beam setting reaches 73.5, against 85.7 for PointRCNN on 64-beam LiDAR
[arxiv:1906.06310 tab. 2]. The four beams are thinned from 64-beam scans, and the
depth ground truth for training still comes from 64-beam LiDAR, so the results do
not show what a real four-beam sensor would give [arxiv:1906.06310 §5.1].

Card: [Pseudo-LiDAR++](/papers/1906.06310).

[MonoDIS](/papers/1905.12365) detects cars from one calibrated RGB image with a
ResNet34 and FPN two-stage design [arxiv:1905.12365 §4.1] [arxiv:1905.12365
§4.2]. Its main change splits the 3D box loss into one copy per parameter group.
In each copy only that group comes from the prediction, and the rest is set to
the ground truth [arxiv:1905.12365 §5.1]. A second head predicts a 3D confidence
from the box error, without extra labels [arxiv:1905.12365 §4.3]. On KITTI test,
car, moderate difficulty, 40-point AP reaches 7.94 with the larger training split
and 4.89 with the standard split, against 5.74 for MonoGRNet on the same metric
[arxiv:1905.12365 tab. 3]. The paper gives no official code [arxiv:1905.12365
§7.4].

[TopoLogic](/papers/2405.14747) predicts which lanes connect at an intersection,
given lanes already detected as ordered point lists. It scores each lane pair
twice: with a learnable function of the gap between one lane's end and the next
lane's start [arxiv:2405.14747 §3.3], and with a sigmoid of the similarity of
their queries [arxiv:2405.14747 §3.4]. The two scores are fused with learnable
weights and passed to a graph layer in the decoder [arxiv:2405.14747 §3.5]
[arxiv:2405.14747 §3.2]. On OpenLane-V2 subset_A with a ResNet-50 backbone, it
reports TOP_ll 23.9 and OLS 44.1 (v2.1.0), against 10.9 and 39.8 for TopoNet
[arxiv:2405.14747 tab. 1]. The geometric term can also be added after training,
which raises TopoNet's TOP_ll from 10.9 to 22.3 [arxiv:2405.14747 tab. 5]. The
released post-processing path uses fixed values in place of the learned
parameters, so it does not reproduce the learned mapping [code:
projects/topologic/models/dense_heads/topologic_head.py#L444-L452]. The authors
say the work is for research only [arxiv:2405.14747 §5].

[A survey of 3D detection for driving](/papers/2106.10823) groups methods by
input: image, point cloud and fusion. It compares published KITTI, nuScenes and
Waymo results [arxiv:2106.10823 tab. 6] [arxiv:2106.10823 tab. 8], and its own
case study runs fifteen detectors on KITTI validation. The case study finds that
3D location error contributes the most error [arxiv:2106.10823 tab. 10]. Its
published tables are not re-measured, so they carry the same caveats as their
source papers.

Cards: [PolarFormer](/papers/2206.15398), [DeepInteraction](/papers/2208.11112),
[Group-Free 3D](/papers/2104.00678), [BEVFusion](/papers/2205.13790),
[AVOD](/papers/1712.02294), [BADet](/papers/2104.10330),
[BEVFormer](/papers/2203.17270), [DETR3D](/papers/2110.06922),
[3D detection survey](/papers/2106.10823),
[OPV2V](/papers/2109.07644), [End-to-end Interpretable Neural Motion Planner](/papers/2101.06679),
[VoTr](/papers/2109.02497), [CaDDN](/papers/2103.01100),
[TopoLogic](/papers/2405.14747), [MonoDIS](/papers/1905.12365),
[Pseudo-LiDAR++](/papers/1906.06310), [CenterFusion](/papers/2011.04841).

### Unknown objects and open-world detection

A closed-set detector labels an unseen object as one of its known classes,
often with high confidence. [Towards Open World Object Detection](/papers/2103.02603)
(ORE) treats unknown objects as something to flag now and learn later. The
detector marks them as unknown, a human names new classes, and the model is
updated without retraining from scratch [arxiv:2103.02603 §3]. ORE is a
Faster R-CNN with three additions [arxiv:2103.02603 §4]:

- a contrastive clustering loss on RoI features, which pulls each class
  towards its own prototype [arxiv:2103.02603 §4.1];
- pseudo-labels for unknowns, taken from high-objectness background boxes that
  do not overlap a ground-truth object [arxiv:2103.02603 §4.2];
- an energy score, with two fitted Weibull distributions that decide whether a
  detection is unknown [arxiv:2103.02603 §4.3].

Incremental steps replay a balanced set of old examples, with 50 per class
chosen on validation [arxiv:2103.02603 §4.4] [arxiv:2103.02603 §5.2].

- On the VOC then COCO task split, ORE lowers Wilderness Impact at task 1 from
  0.0699 to 0.0219 against Faster R-CNN [arxiv:2103.02603 tab. 2].
- At task 2, mAP over both known sets is 38.98 for ORE, against 37.47 for
  Faster R-CNN with finetuning and 14.91 without it [arxiv:2103.02603 tab. 2].
  The no-replay baseline forgets most old classes, so the finetuning row is the
  fairer comparison.
- On the VOC incremental benchmark, ORE has the highest mAP in all three
  splits, 0.3 to 1.4 points above the next-best method [arxiv:2103.02603 tab. 3].

Pseudo-labels come from background boxes, so a real but unlabelled object can
be treated as unknown [arxiv:2103.02603 §4.2]. The released code differs from
the paper in several hyperparameters, including the clustering margin and the
energy temperature; the card lists them [card](/papers/2103.02603).

### Few-shot fine-tuning (FSCE)

[FSCE](/papers/2103.05950) keeps the ResNet backbone frozen and fine-tunes the
FPN, the RPN and the RoI box head on K labelled shots of each novel class, plus
base-class samples [arxiv:2103.05950 §3] [arxiv:2103.05950 fig. 3]. It adds a
small head that encodes each sampled RoI, trained with a supervised contrastive
loss. RoIs of one class are pulled together and the others pushed apart, and the
term enters the total loss with weight 0.5 [arxiv:2103.05950 §3.2]
[arxiv:2103.05950 §3.3] [arxiv:2103.05950 eq. 5]. The authors argue that most
few-shot errors are wrong class labels on boxes that are already well placed
[arxiv:2103.05950 §1].

- Without the contrastive term, RPN and RoI refinement alone reach 49.7 novel
  nAP50 at 3 shots on VOC split 1. The full method reaches 51.4 at 3 shots and
  63.4 at 10 shots [arxiv:2103.05950 tab. 5].
- On COCO, the 30-shot novel AP is 16.4, against 13.7 for TFA w/ cos
  [arxiv:2103.05950 tab. 3].
- The released configs use other λ and IoU thresholds from the paper's text, and
  not every paper cell has a released config [code:
  configs/PASCAL_VOC/split1/3shot_CL_IoU.yml#L29].

Card: [FSCE](/papers/2103.05950).

### Long-tailed classes

When most training boxes belong to a few common classes, a sigmoid classifier
sees far more negative than positive gradient for each rare class.
[EQL v2](/papers/2012.08548) reweights each class's positive and negative
gradients from the ratio of their accumulated magnitudes, in one training stage
[arxiv:2012.08548 §3.2] [arxiv:2012.08548 eq. 5]. On LVIS v1.0 with Mask R-CNN and
ResNet-50 at 1x, it reports 23.7 AP and 14.9 AP on rare classes. The earlier EQL
reports 18.6 and 2.1, softmax 16.1 and 0.0 [arxiv:2012.08548 tab. 1]. The
two-stage BAGS method reports 23.1 AP [arxiv:2012.08548 tab. 1]. On Open Images,
the authors report 52.6 AP against 43.1 for Faster R-CNN, with the LVIS settings
[arxiv:2012.08548 tab. 10]. The EQL row also differs by table: 18.6 AP in tab. 1
and 21.6 in tab. 5, which uses a 2x schedule and the authors' re-implementation
[arxiv:2012.08548 tab. 5].

### Pixel masks for saliency and camouflage

These cards carry the detection tag, but they predict a mask per salient or
camouflaged object rather than one box per object. Their scores are segmentation
scores, so do not put them in a table with box AP. The salient object detection
concept page covers them in more depth.

[Video salient object detection](/papers/1702.00871) runs a static network on
each frame and a dynamic network on frame pairs plus the static map. No optical
flow is computed at test time [arxiv:1702.00871 §3.1] [arxiv:1702.00871 §3.3].
Its training videos are synthetic, made by warping still images
[arxiv:1702.00871 §4]. Training on image-derived video alone raises FBMS MAE
from 7.65% to 9.27% [arxiv:1702.00871 tab. 2], and the authors report about 2 fps
for the full model [arxiv:1702.00871 §5.4].

[DVA](/papers/1705.02544) predicts eye fixations with a VGG16 encoder and three
deconvolution decoders, on conv3-3, conv4-3 and conv5-3, each with its own loss.
A learned fusion layer merges the three maps [arxiv:1705.02544 §3.3.1]
[arxiv:1705.02544 §3.2]. It trains on SALICON only and reports about 0.1 s per
image on one GPU [arxiv:1705.02544 §3.3.3]. It leads on most metrics for four of
the five eye-tracking sets, but not on MIT300, where DeeFix and SALICON score
higher [arxiv:1705.02544 tab. 4].

[EGNet](/papers/1908.08297) adds a salient-edge path built on the conv2-2
features and the top-level location. Its edge features are added to each object
scale before a second prediction, and the four maps are fused [arxiv:1908.08297
§3.2.2] [arxiv:1908.08297 §3.3]. Max F-measure on DUTS-TE is 0.893 with ResNet
and 0.880 with VGG [arxiv:1908.08297 tab. 2]. The paper does not name its
optimiser [arxiv:1908.08297 §4.1].

[D3Net](/papers/1907.06781) runs three streams: RGB only, RGB with depth, and
depth only. At test time it keeps the RGB-plus-depth map only when it agrees with
the depth-only map, within a mean absolute difference of 0.15, and otherwise
uses the RGB map [arxiv:1907.06781 §4.2] [arxiv:1907.06781 eq. 3]. On NJU2K test
its S-measure is 0.900, against 0.879 for CPFP [arxiv:1907.06781 tab. 4]. The
0.15 threshold was chosen on data the paper does not name [arxiv:1907.06781
§4.3].

[SINet-V2](/papers/2102.10274) finds camouflaged objects in two stages. A
neighbour connection decoder first gives a coarse map, and three group-reversal
attention blocks then refine it [arxiv:2102.10274 §4.2] [arxiv:2102.10274 §4.3].
On the COD10K test set its S-measure is 0.815, on CAMO 0.820 and on CHAMELEON
0.888 [arxiv:2102.10274 tab. 2]. Its dataset, COD10K, has 10,000 images
[arxiv:2102.10274 §3.1].

[Anabranch Network](/papers/2105.09451) multiplies one scalar into the map of any
saliency network. The scalar is the probability that camouflage is present in
the image, so an image without camouflage is pushed towards an empty map
[arxiv:2105.09451 §4.2]. On CAMO-COCO, ANet-SRM reaches F-beta 0.826 and IoU
0.732, against 0.699 and 0.535 for the pre-trained SRM [arxiv:2105.09451 tab. 3].

[COD-Rank](/papers/2103.04011) predicts a fixation map, a camouflage mask and a
difficulty rank in one network. The ranks come from eye-tracker detection delays
[arxiv:2103.04011 §3.1] [arxiv:2103.04011 §3.2] [arxiv:2103.04011 §3.3]. Its
joint model trained on COD10K and CAMO reaches an S-measure of 0.793 on COD10K,
against 0.776 for SINet [arxiv:2103.04011 tab. 4]. The released code has no
ranking branch, so the rank results cannot be checked from it [code:
joint_cod_col/train.py].

Cards: [Video salient object detection](/papers/1702.00871),
[DVA](/papers/1705.02544), [EGNet](/papers/1908.08297),
[D3Net](/papers/1907.06781), [SINet-V2](/papers/2102.10274),
[COD-Rank](/papers/2103.04011), [Anabranch Network](/papers/2105.09451).

### Surveys

[A Survey of Modern Deep Learning based Object Detection Models](/papers/2104.11892)
covers datasets and metrics, backbones, two-stage and single-stage detectors, and
lightweight networks. Its detector table collects author-reported COCO or VOC AP
and FPS for 17 detectors, without re-running any of them [arxiv:2104.11892 §1]
[arxiv:2104.11892 tab. 3]. Its summary is that two-stage detectors are generally
more accurate and single-stage detectors are faster [arxiv:2104.11892 §9]. Its
headline rows are DetectoRS at 53.3 AP on COCO with a 1333 input, at about 4 FPS,
and YOLOv4 at 43.0 AP with a 512 input, at about 31 FPS [arxiv:2104.11892 tab. 3].
The rows differ in hardware, input size and training data, so the survey calls
them a guide rather than a ranking [arxiv:2104.11892 §7]. Its FPS column mixes
devices, so measure speed on your own hardware.

Card: [A Survey of Modern Deep Learning based Object Detection Models](/papers/2104.11892).

[Image Data Augmentation for Deep Learning: A Survey](/papers/2204.08610) groups
augmentation into manipulation, erasing, mixing, automated policy search,
feature-space and GAN methods [arxiv:2204.08610 §2] [arxiv:2204.08610 §3]. Its own
experiments use only the first three groups, on segmentation, classification and
detection [arxiv:2204.08610 §4]. For detection on COCO2017, Faster R-CNN rises from
36.40 to 36.80 mAP with augmentation, and CenterNet falls from 41.42 to 41.15
[arxiv:2204.08610 tab. 4]. No training settings or run-to-run variance are
reported [arxiv:2204.08610 §4.3].

[Moving Forward](/papers/2411.10291) is a survey of autonomous-driving sensors,
software and hardware. Its object detection table lists Faster R-CNN, YOLO, SSD,
VoxelNet, PointPillar and other 2D and 3D detectors, with mAP and FPS taken from
the original works and not re-measured by the authors [arxiv:2411.10291 tab. II].
The table mixes VOC2007 2D results with KITTI 3D results, and it does not say
what hardware the FPS came from, so it is a reference, not a ranking
[arxiv:2411.10291 tab. II]. The authors' own measurements cover end-to-end driving
models, not detection alone [arxiv:2411.10291 §V-A].

[A Review on Deep Learning in UAV Remote Sensing](/papers/2101.10861) screened
232 articles and proceedings from 2016 to 2021 and tallies them by network, task,
sensor and setting. Object detection is 53.9% of the coded studies and
segmentation 40.7%. The authors' tallies are not independent measurements
[arxiv:2101.10861 §3]. Detection studies mostly use RGB imagery
[arxiv:2101.10861 §3].

[A Survey of Deep Learning Techniques for Weed Detection from Images](/papers/2103.01415)
reviews 70 deep-learning papers from 2016 to August 2020 on finding weeds in
crop images. Its detector list covers Faster R-CNN, SSD, Mask R-CNN and YOLO-v3
and tiny YOLO-v3 [arxiv:2103.01415 §10.2]. Most of the 19 public datasets are
specific to one site or crop [arxiv:2103.01415 §6.3]. The reported scores come
from different crops, metrics and datasets, and the survey says they are hard to
compare [arxiv:2103.01415 §11].

[3D Object Detection for Autonomous Driving: A Comprehensive Survey](/papers/2206.09474)
organises 3D detection by sensor: LiDAR, camera and multi-modal
[arxiv:2206.09474 §3] [arxiv:2206.09474 §4] [arxiv:2206.09474 §5]. It concludes
that LiDAR and fusion methods are the choice when speed and accuracy matter most,
and camera methods when cost matters most [arxiv:2206.09474 §10.1.6]. Its
tables copy scores from original papers that use different KITTI, nuScenes and
Waymo splits, so they are not one ranking [arxiv:2206.09474 tab. 16]
[arxiv:2206.09474 tab. 17].

### Benchmarks and datasets

Some cards are datasets or competitions rather than methods. Their baselines
give a starting point and a metric, not a ranking of methods.

[xView](/papers/1802.07856) is an overhead dataset with over one million boxes in
60 classes. Its SSD baseline reaches total mAP 0.1456 on 300-pixel chips and
0.2590 with chips of three sizes [arxiv:1802.07856 tab. 1]. Augmentation on top
of the three sizes gives 0.1549, so it does not help in this setup
[arxiv:1802.07856 §4].

[The UAV maritime benchmark](/papers/2105.01922) provides detection and tracking
labels for swimmers, life jackets and boats, with per-frame altitude and camera
angle. The best detector in its baselines is Faster R-CNN with ResNeXt-101-FPN at
AP 30.4, and EfficientDet-D0 reaches 20.8 [arxiv:2105.01922 tab. 4]. The test
labels are withheld and scored on a server [arxiv:2105.01922 §3.3].

[The ICDAR 2019 receipt competition](/papers/2103.10213) includes a text-line
localisation task. The top entry reaches Hmean 98.59% [arxiv:2103.10213 §III-A2]
[arxiv:2103.10213 fig. 3]. This is detection of text rather than objects, so its
numbers are not comparable with the box-detection results above.

[ARKitScenes](/papers/2111.08897) records about 5,000 RGB-D captures of 1,661
indoor scenes with a tablet's LiDAR sensor, with oriented 3D boxes for 17
furniture classes [arxiv:2111.08897 §3]. Three indoor detectors trained on it
reach whole-scene mAP of 0.358 (VoteNet), 0.383 (H3DNet) and 0.419 (MLCVNet) at
IoU 0.25 [arxiv:2111.08897 tab. 2]. The per-frame setting changes the order, with
VoteNet first [arxiv:2111.08897 tab. 3]. All numbers are author-reported.

[RDD2022](/papers/2209.08538) is a road-damage dataset of 47,420 photos from six
countries, with 55,007 labelled boxes in the training split for four classes:
longitudinal and transverse cracks, alligator cracks and potholes
[arxiv:2209.08538 tab. 2] [arxiv:2209.08538 fig. 5]. Test labels are withheld and
scored through a leaderboard [arxiv:2209.08538 p. 11]. The paper reports dataset
counts only, with no detector results [arxiv:2209.08538 p. 1].

[Synscapes](/papers/1810.08705) is a synthetic street-scene set of 25,000
path-traced images, with instance maps and per-object 2D and 3D boxes
[arxiv:1810.08705 §3]. Faster R-CNN trained on it alone reaches 0.488 mAP on its
own validation split [arxiv:1810.08705 tab. 6]. Adding it to KITTI training
raises KITTI-validation mAP from 0.456 to 0.519 [arxiv:1810.08705 tab. 6]. The
paper does not state the dataset's licence or download location, so check its
terms before reuse [arxiv:1810.08705 §3].

Cards: [xView](/papers/1802.07856), [Synscapes](/papers/1810.08705),
[UAV maritime benchmark](/papers/2105.01922),
[ICDAR 2019 receipts](/papers/2103.10213), [ARKitScenes](/papers/2111.08897),
[RDD2022](/papers/2209.08538).

### Detection as a component of other tasks

Some cards use detection as one step in a larger system. Their headline results
measure the larger task, not the detector.

[POI](/papers/1610.06136) tracks people with a Faster R-CNN detector and a
128-dimensional appearance embedding trained with softmax and triplet losses
[arxiv:1610.06136 §1] [arxiv:1610.06136 §2]. The authors show that detection
errors carry into tracking. Adding skip pooling and multi-region features cut
false positives plus false negatives on MOT16 train from 52,727 to 46,587
[arxiv:1610.06136 tab. 1]. On the MOT16 challenge, the online tracker reaches
MOTA 66.1 and the offline one 68.2, both author-reported [arxiv:1610.06136 tab. 4].

[YOLOP](/papers/2108.11250) runs one shared encoder with three heads, for
vehicles, drivable area and lane lines [arxiv:2108.11250 §3]. Only vehicles are
scored for detection [arxiv:2108.11250 §4.2.1]. Its jointly trained model reaches
76.5 vehicle mAP50 at 41 fps on one TITAN XP [arxiv:2108.11250 tab. 1]. One
multi-task pass takes 24.4 ms, against about 45 ms for the three single-task
models summed from the same table [arxiv:2108.11250 tab. 5].

[YOLO-Pose](/papers/2204.06806) adds 51 keypoint values per anchor to a
YOLOv5-style head. Each person's box and 17 keypoints then come from one pass,
with no grouping step [arxiv:2204.06806 §3] [arxiv:2204.06806 eq. 1]. The keypoint
loss is one minus the object keypoint similarity [arxiv:2204.06806 §3.4]
[arxiv:2204.06806 eq. 3]. The largest model reaches 90.2 AP50 on val2017, but its
AP there is 69.4, below 69.9 for HigherHRNet-W48 at 640 [arxiv:2204.06806 tab. 1].

[MARTENS](/papers/2410.19277) places an oriented-box YOLOv8 detector inside a
simulated pick-and-place cell. A genetic search over box poses and lighting looks
for scenes where the arm fails, and the detector is then fine-tuned on the failing
images [arxiv:2410.19277 §2.2] [arxiv:2410.19277 §2.5]. Author-reported: the
search finds more failures than random search in both use cases [arxiv:2410.19277
tab. 2] [arxiv:2410.19277 tab. 3], and fine-tuning repairs more than 99% of them
[arxiv:2410.19277 tab. 6]. A detector that starts from the repaired simulation
model reaches 0.948 and 0.892 mAP after training on about 100 real images per
use case. A real-only detector reaches 0.806 and 0.755 [arxiv:2410.19277 tab. 8].
Each simulated test takes 280 to 310 s on one GPU, so the search is costly
[arxiv:2410.19277 §3.1].

[The agricultural deep-learning survey](/papers/1807.11809) tabulates 40 studies.
Weed identification and fruit counting are among its application areas, with
five and four studies respectively [arxiv:1807.11809 §4.1]. Most of its reported
numbers are classification accuracy, not detection, and setups differ across
studies, so they are not directly comparable [arxiv:1807.11809 §4.9].

[Florence-2](/papers/2311.06242) handles detection, grounding and captioning in
one prompt-driven model, with boxes written as location tokens
[arxiv:2311.06242 §3]. Its detection results are reported as one task among
many [arxiv:2311.06242 tab. 4].

[The scene graph survey](/papers/2104.01111) covers the step after detection,
where objects are linked by relations. Its advice is to report mean recall per
predicate, because plain recall is dominated by common relations
[arxiv:2104.01111 §6.2].

[The autonomous racing survey](/papers/2202.07008) lists high-speed object
detection and localisation without HD maps as open problems
[arxiv:2202.07008 §2.1]. [The remote sensing transformer survey](/papers/2209.01206)
reproduces detection numbers from the original papers, such as DOTA mAP up to
80.14 [arxiv:2209.01206 §5]. Neither card re-runs its numbers.

[Visual-RFT](/papers/2503.01785) fine-tunes a vision-language model with GRPO,
using a reward that combines the IoU of each predicted box with its matching
ground-truth box, plus a confidence term [arxiv:2503.01785 §3.2.1] [arxiv:2503.01785
eq. 5]. On 8 COCO categories with two images each, the 2B model reaches 41.5 mAP,
against 21.0 for supervised fine-tuning and 19.6 for the base model
[arxiv:2503.01785 tab. 3]. The paper does not state the IoU threshold used in the
reward [arxiv:2503.01785 §3.2.1].

[The anomaly detection survey for driving](/papers/2204.07974) lists lidar
object-level anomaly detection as only just gaining momentum
[arxiv:2204.07974 §3]. It runs no experiments of its own.

[The social-distancing monitor](/papers/2005.01385) runs a YOLO v3 person detector
with DeepSORT tracking, then groups people who stand close together
[arxiv:2005.01385 §IV-A]. On the authors' 800-image person set, YOLO v3 reaches
0.846 mAP at 23 FPS, against 0.969 for Faster R-CNN at 3 FPS and 0.691 for SSD at
10 FPS [arxiv:2005.01385 tab. III]. The authors call YOLO v3 balanced across mAP,
training time and speed [arxiv:2005.01385 §V]. The paper does not state the IoU
threshold behind its mAP, and it never checks the grouping or the violation index
against ground truth [arxiv:2005.01385 §VI].

[SportsMOT](/papers/2304.05170) is a tracking dataset for players in basketball,
volleyball and football. Its trackers take per-frame boxes from a YOLOX detector
[arxiv:2304.05170 §5.1]. The authors report that the Kalman motion model lowers
HOTA on this data, 64.1 with the filter against 71.5 without it [arxiv:2304.05170
tab. 6].

[GenEval](/papers/2310.11513) uses a detector as a checker for text-to-image
output. A Mask2Former with a Swin-S backbone, trained on the 80 COCO classes,
finds the objects in each generated image, and a confidence threshold of 0.3
decides which detections count, or 0.9 for counting [arxiv:2310.11513 §3.2]
[arxiv:2310.11513 app. C.3]. The authors state that COCO-trained detectors cannot
verify fine details or artistic renderings [arxiv:2310.11513 §6]. On the paper's
six-task overall score, IF-XL reaches 0.61, with position at 0.13 and attribute
binding at 0.35 [arxiv:2310.11513 tab. 2].

Cards: [POI](/papers/1610.06136), [YOLOP](/papers/2108.11250),
[YOLO-Pose](/papers/2204.06806), [Social-distancing monitor](/papers/2005.01385), [SportsMOT](/papers/2304.05170),
[Agricultural deep-learning survey](/papers/1807.11809),
[Florence-2](/papers/2311.06242),
[scene graph survey](/papers/2104.01111),
[autonomous racing survey](/papers/2202.07008),
[remote sensing transformer survey](/papers/2209.01206),
[Visual-RFT](/papers/2503.01785),
[anomaly detection survey](/papers/2204.07974),
[GenEval](/papers/2310.11513).

### Anchor-free boxes on a timeline

Temporal action localisation finds the start and end of an action in a video, so
it is detection along time. [AFSD](/papers/2103.13137) is anchor-free: a 1D pyramid
predicts the distance to the start and end of an action at each time step, and a
boundary-pooling refinement takes the most salient boundary features of each
proposal [arxiv:2103.13137 §3]. Author-reported, it reaches 52.0 average mAP on
THUMOS14 with I3D features and 34.4 on ActivityNet 1.3 [arxiv:2103.13137 tab. 1].
The table mixes backbones (I3D, TS, C3D, P3D), so compare rows within one backbone.

### Detection and density for crowd counting

[DecideNet](/papers/1712.06679) counts people by blending two maps: a detector's
head boxes turned into Gaussian bumps, and a five-layer regression map. A small
attention network picks the blend pixel by pixel [arxiv:1712.06679 §3.4]
[arxiv:1712.06679 eq. 4]. The authors report the lowest MAE of the compared
methods on Mall, 1.52 against 1.82 for Exemplary Density [arxiv:1712.06679
tab. 1], and an average MAE of 9.23 on WorldExpo'10 [arxiv:1712.06679 tab. 3].
On ShanghaiTech PartB it reports 21.53, while CP-CNN reports 20.1
[arxiv:1712.06679 tab. 2]. In the authors' ablation, the regression map alone
gives 3.37 MAE on Mall and the detection map alone 4.50 [arxiv:1712.06679 tab. 4].
The Gaussian width of the detector map and the loss weight λ are not stated
[arxiv:1712.06679 §3.3] [arxiv:1712.06679 §4]. Neither training nor inference
cost is reported [arxiv:1712.06679 §5.1].

Card: [DecideNet](/papers/1712.06679).

## Evaluation

The Metrics Reloaded consensus gives detection-specific rules for choosing a
score [arxiv:2206.01653 §ObD]:

- Accuracy, specificity and AUROC are not valid for detection, because there is
  no meaningful count of true negatives. Use the three remaining entries of the
  confusion matrix [arxiv:2206.01653 §ObD].
- With no continuous class scores, a counting metric such as F-beta is the
  default. With scores, the paper pairs a counting metric with AP or FROC
  [arxiv:2206.01653 §ObD].
- The localisation criterion and the assignment strategy are chosen by decision
  guides in the supplement, which this page does not transcribe
  [arxiv:2206.01653 §ObD] [supplement: decision guides].

The paper reports no benchmark numbers. Its rules come from a multi-round expert
consensus, with each component above 75% agreement [arxiv:2206.01653 §Step 2].

For 3D detection on KITTI, the cards mix two AP conventions. Test-server results
use 40 recall positions, and the validation tables use 11 [arxiv:2104.10330 §4.2]
[arxiv:2104.10330 tab. 2]. The 3D survey's case study also uses 11 positions on
validation, while its headline test table uses 40 [arxiv:2106.10823 §4.2.1]
[arxiv:2106.10823 tab. 6]. Numbers from the two conventions should not be placed
in one table.

The KITTI 11-point AP has a flaw that inflates scores at the low end. Precision
at recall 0 is counted, so one correct detection gives 1/11, about 9.09% AP, per
difficulty level [arxiv:1905.12365 §6] [arxiv:1905.12365 tab. 5]. The
[MonoDIS](/papers/1905.12365) authors switch to 40 recall points, which drops
recall 0 [arxiv:1905.12365 §6]. Their test AP40 values were computed from
published precision-recall values, not from the server [arxiv:1905.12365 §7.4].
Use AP40 only with other AP40 numbers.

For masks, [Structure-measure](/papers/1708.00786) scores a non-binary map as
an average of two terms with equal weight, `S = 0.5·S_o + 0.5·S_r`
[arxiv:1708.00786 §4.3]. The region term compares four blocks, split at the
ground-truth centroid, with SSIM. The object term compares the mean and spread
of the map inside and outside the object [arxiv:1708.00786 §4.1]
[arxiv:1708.00786 §4.2]. In the authors' user study, 63.69% of trials preferred
the top map by this measure over the one from weighted F-measure
[arxiv:1708.00786 §5.6]. The official code weights the blocks by image area, not
by ground-truth area as the paper describes, so a port has to state its choice
[code: S_region.m#L84-L88].

Open-world detection uses two error measures next to known-class mAP. Wilderness
Impact (WI) and Absolute Open-Set Error (A-OSE) both count how often unknown
objects are predicted as known classes, and lower is better [arxiv:2103.02603 §5.1].
WI cannot be measured on the last task, because no unknown boxes remain
[arxiv:2103.02603 §5.3].

## When to use which

| Situation | Use | Why |
|---|---|---|
| Must run on a phone or a small CPU | Single-stage SSD with MobileNet | Fastest measured option in the road-damage paper: 30.6 ms on GPU against 63.1 ms for Inception V2 [arxiv:1801.09454 tab. 3]. Phone timing is on old hardware and an older framework |
| Microcontroller with 1 to 2 MB Flash and under 1 MB RAM, no accelerator, low frame rate acceptable | microYOLO, with pruning and 8-bit quantization | 3.46 FPS at 128 x 128 on OpenMV H7 R2, 324 KB RAM [arxiv:2408.15865 tab. 2]. Accuracy is modest and task-dependent [arxiv:2408.15865 §3.1] |
| Real-time on a laptop CPU with no GPU, accuracy secondary | YOLO-LITE | 21 FPS at 33.57% VOC mAP, against 2.4 FPS and 40.48% for Tiny-YOLOv2 [arxiv:1811.05588 tab. III]. One device, author-reported; COCO mAP is 12.26% [arxiv:1811.05588 tab. V] |
| Fixed small set of classes, video rate on a GPU | Cascade of small single-stage detectors | 47 FPS for a four-stage pipeline on Titan Xp [arxiv:1802.09567 tab. 5]. Only works when the layout is fixed; the cascade is built for one plate format [arxiv:1802.09567 §4.3] |
| Single-stage, but small objects need fused scales | FSSD on SSD | Fuses three layers before the heads; 78.8 mAP on VOC2007 at 65.8 FPS on a 1080Ti, author-reported [arxiv:1712.00960 tab. 3]. Speed cost is small but not quantified in the card [arxiv:1712.00960 §4] |
| Tiny objects that few anchors or points overlap | Two-stage detector with RFLA label assignment | Faster R-CNN from 11.1 to 21.1 AP on AI-TOD [arxiv:2208.08738 tab. 5]. Drop-in change to positive selection; the card records no extra inference cost [arxiv:2208.08738 §1] |
| Regions that are not objects, or extra input views | Two-stage detector with the extra view fused after RoI pooling | Fusion beats either stream alone on three of four forgery benchmarks, not on Columbia [arxiv:1805.04953 §4.2]. Budget for roughly double the backbone compute [arxiv:1805.04953 §3.2] |
| Need a box per object at high accuracy on a medical benchmark | Single-stage YOLOv4 or RetinaNet | YOLOv4 at AP 0.8513 and 48 FPS; RetinaNet R101 at 0.8745 and 16.8 FPS [arxiv:2011.07631 tab. III]. The authors favour YOLOv4 for speed [arxiv:2011.07631 §VI] |
| Few labelled boxes, and a vision-language model is available | RL fine-tuning with a box reward (Visual-RFT) | 41.5 mAP at two shots against 21.0 for SFT on 8 COCO categories, 2B model [arxiv:2503.01785 tab. 3]. The reward's IoU threshold is not stated [arxiv:2503.01785 §3.2.1]; the paper does not report GPU cost [arxiv:2503.01785 §4] |
| Lots of unlabelled images, few boxes | Teacher-student training with pseudo-labels | 20.75 mAP at 1% COCO labels against 13.97 for STAC [arxiv:2102.09480 tab. 1]. Needs a burn-in stage and a second training loop [arxiv:2102.09480 §3] |
| Tilted objects in aerial or satellite images, end-to-end set prediction | OrientedFormer | 75.37 AP50 on DOTA-v1.0 with ResNet-50 against 74.16 for ARS-DETR, at 12 epochs against 36 [arxiv:2409.19648 tab. 4] [arxiv:2409.19648 §4.8]. Needs about 2.8 times the FLOPs [arxiv:2409.19648 tab. 14]. Oriented R-CNN is higher on the same single-scale DOTA result, at 75.87 [arxiv:2409.19648 tab. 4] |
| Want no NMS at inference and can afford long training | DETR-style set prediction, optionally with auxiliary heads | The base method runs unchanged at inference [arxiv:2211.12860 §1]. Auxiliary heads add training cost: 70 to 120 GPU hours for two heads [arxiv:2211.12860 tab. 7] |
| Want fast DETR-style convergence | TSP with a transformer encoder, or H-DETR training | TSP reaches 43.1 AP in 36 epochs against 500 for DETR [arxiv:2011.10881 tab. 1]. H-DETR adds 1.7 AP for 12 epochs with no inference cost [arxiv:2207.13080 tab. 1] |
| Training budget tight | Plain detector without auxiliary heads | One head costs 1.2 times the baseline GPU hours and two cost 1.7 times, for 1.6 to 5.8 AP across the reported settings [arxiv:2211.12860 tab. 2] [arxiv:2211.12860 tab. 7] |
| Rare classes in the training set | Expect low recall, and rebalance the data | Low recall on the rarest road-damage classes [arxiv:1801.09454 §5] |
| Indoor 3D scenes with no LiDAR | Group-free transformer detector | 67.3 mAP@0.25 on ScanNet V2 for the best run [arxiv:2104.00678 tab. 1]. Needs point clouds, not images |
| Camera and LiDAR, LiDAR may fail | BEV fusion with a separate camera stream | 67.9 against 64.9 mAP for LiDAR alone on nuScenes validation [arxiv:2205.13790 tab. 1]. Abstract reports larger gains under simulated LiDAR loss [arxiv:2205.13790 §4.4.1] |
| Several vehicles sharing sensor data over a radio link | Intermediate fusion with compressed features | Keeps most AP under compression and fits the link budget in the authors' estimate [arxiv:2109.07644 §5.6]. Simulated data only |
| Small objects in a real-time single-stage detector, one dataset | YOLO-Z: a higher-resolution map in place of the lowest | Small-object mAP@0.5 from 0.869 to 0.925 at the S scale, for +0.9 ms on a GTX 1080 Ti [arxiv:2112.11798 tab. 2]. One cone dataset, single runs |
| One-stage training, better positives, no second stage | DDOD on ATSS, RetinaNet or FCOS | 39.4 to 41.6 AP on ATSS [arxiv:2107.02963 tab. 4]. Replaces centerness with an IoU branch; no matched speed comparison [arxiv:2107.02963 tab. 8] |
| Many classes with few examples each, sigmoid head, one stage | EQL v2 gradient reweighting | 23.7 AP and 14.9 rare-class AP on LVIS v1.0, against 23.1 AP for the two-stage BAGS [arxiv:2012.08548 tab. 1] |
| Person detection on a fixed camera with one GPU | YOLO v3, not Faster R-CNN | 0.846 mAP at 23 FPS, against 0.969 at 3 FPS for Faster R-CNN, on the authors' 800-image set [arxiv:2005.01385 tab. III]. The distance rule is not validated |
| LiDAR plus cameras, accuracy first, a few FPS is enough | DeepInteraction | 70.8 mAP on nuScenes test, against 65.5 for LiDAR-only TransFusion-L [arxiv:2208.11112 tab. 1]. 2.6 FPS on a V100, against 3.8 for TransFusion [arxiv:2208.11112 tab. 2]. Needs calibration [arxiv:2208.11112 app. A.3] |
| Radar and camera on one sensor set, velocity needed from a single frame | CenterFusion | NDS 0.453 against 0.328 for the camera baseline on nuScenes val, mAP 0.332 against 0.306 [arxiv:2011.04841 tab. 1]. Test mAP is below the Hourglass CenterNet, 0.326 against 0.338 [arxiv:2011.04841 tab. 1] |
| Surround cameras only, polar-grid layout | PolarFormer | 39.6 mAP on nuScenes validation with ResNet-101 [arxiv:2206.15398 tab. 3]. No runtime reported. The 50.0 setting needs extra pretraining data |
| Synthetic street scenes with exact labels | Synscapes, mixed with KITTI | KITTI-validation mAP from 0.456 to 0.519 when Synscapes is added [arxiv:1810.08705 tab. 6]. Licence not stated; do not redistribute until it is confirmed |
| Table structure in document images | DETR trained on PubTables-1M | 0.8138 content accuracy on canonical test, against 0.1039 for Faster R-CNN [arxiv:2110.00061 tab. 4]. Scores depend on the test split |
| Real-time detector on a TensorRT GPU, and no NMS step in the serving code | YOLOv10 | 51.1 COCO val AP at 4.74 ms for M, against 50.6 at 9.50 ms for YOLOv8-M [arxiv:2405.14458 tab. 1]. Latency is FP16 on a T4 with batch 1 and excludes I/O [arxiv:2405.14458 §4.1] [arxiv:2405.14458 app. A.1]. Latency is only comparable in an exported format [code: README.md#L55] |
| Tilted objects in aerial images, small model | ReDet | 80.10 mAP on DOTA-v1.0 with multi-scale training, against 79.42 for the best other row in its table, with a 12 Mb backbone against 103 Mb [arxiv:2103.07733 tab. 6] [arxiv:2103.07733 tab. 1]. The abstract's gains do not match these rows, so use the tables [arxiv:2103.07733 §1] |
| Plug a stronger head into an existing one-stage detector | Dynamic Head on ATSS | 39.4 to 42.6 COCO val AP at ResNet-50, 1x [arxiv:2106.08322 tab. 3]. Six blocks add 21.5 GFLOPs to a 254.39 GFLOP baseline [arxiv:2106.08322 tab. 2], so the abstract's "without any computational overhead" is not literal [arxiv:2106.08322 §4.2] |
| Dense detector with a transformer backbone | Swin-T | 50.5 box AP on COCO val with Cascade Mask R-CNN, against 46.3 for ResNet-50 [arxiv:2103.14030 tab. 2] |
| Dense detector with a transformer backbone, and compute matters more than the last point of AP | PVT v2-B2-Li | ATSS: 48.9 box AP at 194 GFLOPs, against 47.2 AP at 215 GFLOPs for Swin-T [arxiv:2106.13797 tab. 4]. The linear variant loses 1.0 RetinaNet AP against the convolutional-SRA model [arxiv:2106.13797 tab. 6] |
| Cheap accuracy gain on a YOLO model, and memory has headroom | RFAConv in place of the 3×3 convolutions | YOLOv8n COCO2017 AP 36.4 to 37.7, with per-image validation 4.2 to 4.5 ms [arxiv:2304.03198 tab. 7]. The memory cost grows by k² in the expanded tensor [arxiv:2304.03198 §4.6]; the authors limit the number of replacements when speed matters [arxiv:2304.03198 §4.3] |
| Hybrid CNN and transformer backbone, and a small accuracy gain over ResNet-101 is worth some speed | Conformer-S/32 | 43.1 box AP against 39.4 for ResNet-101 FPN in Faster R-CNN [arxiv:2105.03889 tab. 3]. 13.5 against 15.9 FPS on a V100 [arxiv:2105.03889 app. C] |
| Surround cameras, no LiDAR, previous frames available | BEVFormer | 51.7 NDS on nuScenes val with ResNet-101, against 44.8 for its spatial-only variant [arxiv:2203.17270 tab. 2]. Temporal self-attention reads the previous BEV map [arxiv:2203.17270 §3.4] |
| Surround cameras, no depth network, no NMS | DETR3D | 0.374 NDS and 0.303 mAP on nuScenes val with ResNet-101, and 0.434 NDS with CBGS [arxiv:2110.06922 tab. 1]. Its mATE is worse than FCOS3D's [arxiv:2110.06922 §4.2] |
| LiDAR only, car detection on KITTI, accuracy first | BADet | 81.61 AP3D moderate on KITTI test, level with Voxel R-CNN at 81.62 [arxiv:2104.10330 tab. 1]. 140 ms average on a GTX 1080 Ti [arxiv:2104.10330 §4.5] |
| Camera only at run time, car boxes in driving scenes, LiDAR available for training | CaDDN | 13.41 AP R40 moderate on KITTI test, against 9.71 for M3D-RPN [arxiv:2103.01100 tab. 1]. Training needs LiDAR-derived depth maps [arxiv:2103.01100 §3.4]; the released code loads precomputed maps, see [the card](/papers/2103.01100) |
| Stereo cameras at run time, a few LiDAR beams available, 3D car boxes | Pseudo-LiDAR++ with the graph depth correction | 76.9 AP_BEV on KITTI val, moderate, IoU 0.7, with simulated four-beam LiDAR, against 56.0 for the pseudo-LiDAR baseline [arxiv:1906.06310 tab. 1]. Depth ground truth for training still needs 64-beam LiDAR [arxiv:1906.06310 §5.1]; the released code has no detector |
| LiDAR-only voxel detector where wider context helps, and a lower frame rate is acceptable | VoTr-SSD | 78.25 AP3D moderate on KITTI test, against 75.96 for SECOND [arxiv:2109.02497 tab. 2]. 14.65 Hz against 20.73 Hz for SECOND [arxiv:2109.02497 tab. 8]. Waymo LEVEL_1 mAP 68.99 against 67.94 for the authors' SECOND re-run [arxiv:2109.02497 tab. 1] |
| Camera and LiDAR, 3D car boxes with orientation, about 10 frames per second | AVOD | 71.88 AP3D moderate and 66.38 hard on KITTI test, at about 0.1 s per frame on a TITAN Xp [arxiv:1712.02294 tab. 2] [arxiv:1712.02294 §4]. The results are from 2017 and author-reported |
| Lane connectivity at intersections, scored on OpenLane-V2 | TopoLogic (geometric gap term plus query similarity) | 23.9 TOP_ll and 44.1 OLS against 10.9 and 39.8 for TopoNet, author-reported [arxiv:2405.14747 tab. 1]. The geometric term alone gives 20.1 TOP_ll, against 12.9 for similarity alone [arxiv:2405.14747 tab. 4] |
| Known classes plus a stream of unknown objects that must be flagged and learned later | ORE on Faster R-CNN | Task 1 WI 0.0219 against 0.0699 for Faster R-CNN; task-2 mAP over both known sets 38.98 against 37.47 with finetuning [arxiv:2103.02603 tab. 2]. Adds about 0.13 s per training iteration and a replay memory of about 34 MB at 50 exemplars per class [arxiv:2103.02603 app. H] |
| Existing CIoU or SIoU box loss, no architecture change, ratio tuned on your own data | Inner-IoU | Gains of 0.61 to 1.21 mAP50:95 on VOC with YOLOv7-tiny [arxiv:2311.02877 tab. 1]. On AI-TOD with YOLOv5s the change is mixed, from −0.17 to +0.17 [arxiv:2311.02877 tab. 2]. The paper gives no ratio-selection rule and single runs [arxiv:2311.02877 §4.2] |
| Real-time GPU detector, one-stage, no second stage | RTMDet-s for speed-accuracy, RTMDet-x for accuracy | RTMDet-s 44.6 AP at 1.22 ms and RTMDet-x 52.8 AP at 3.10 ms, on the authors' FP16 TensorRT setup with NMS excluded [arxiv:2212.07784 tab. 2] [arxiv:2212.07784 §4.1]. Latency is not comparable across different hardware or input handling |
| One model for detection, captioning and grounding, switched by instruction | VisionLLM with ResNet-50 | 44.6 box AP on COCO against 44.8 for the separately trained model [arxiv:2305.11175 tab. 1]. Parallel decoding is not measured for speed [arxiv:2305.11175 §4.2]. Use InternImage-H, 60.2 AP, when accuracy matters more than the backbone cost [arxiv:2305.11175 tab. 1] |
| Few labelled shots of a new class, with a base-class detector available | FSCE fine-tuning with a contrastive proposal head | 51.4 nAP50 at 3 shots and 63.4 at 10 shots on VOC split 1 [arxiv:2103.05950 tab. 5]. Against TFA w/ fc, the best unstarred row in the split-3 column, the 10-shot gap is 8.3 points [arxiv:2103.05950 tab. 2]. Released configs use other λ and IoU values than the text |
| Monocular RGB only, car boxes scored with KITTI-style AP | MonoDIS, with the larger training split | 7.94 moderate AP40 on KITTI test, against 5.74 for MonoGRNet [arxiv:1905.12365 tab. 3]. Needs camera intrinsics at train and test time [arxiv:1905.12365 §3] |
| One pass for person boxes and keypoints, no grouping step | YOLO-Pose on a YOLOv5-style head | 90.2 AP50 for the largest model on val2017 [arxiv:2204.06806 tab. 1]. Its AP there is 69.4, below 69.9 for HigherHRNet-W48 at 640, so compare AP50 with AP50 only |
| One network for vehicles, drivable area and lanes on embedded hardware | YOLOP, jointly trained | 76.5 vehicle mAP50 at 41 fps on a TITAN XP, and 91.5 drivable mIoU [arxiv:2108.11250 tab. 1] [arxiv:2108.11250 tab. 2]. One multi-task pass takes 24.4 ms, against about 45 ms for three single-task models [arxiv:2108.11250 tab. 5] |
| Camouflaged object masks on one GPU, no box output needed | SINet-V2 | COD10K S-measure 0.815 and CAMO 0.820 [arxiv:2102.10274 tab. 2]. About 45 fps on one GPU, excluding I/O [arxiv:2102.10274 §4.4.2] |
| RGB-D salient object masks where the depth map may be poor | D3Net with the depth gate | S-measure 0.900 on NJU2K test, against 0.879 for CPFP [arxiv:1907.06781 tab. 4]. CPFP leads on LFSD [arxiv:1907.06781 tab. 4]. The gate threshold was picked on data the paper does not name [arxiv:1907.06781 §4.3] |
| Vehicle with a Jetson-class board, cloud retrains the weights, and a single-stage detector is enough | Edge YOLO; use TensorRT FP16 only if latency is the priority | 0.473 mAP@0.5 at 26.6 FPS on Xavier, against 0.657 at 2.1 FPS for YOLOv4 [arxiv:2205.14942 tab. 5]. TensorRT FP16 gives 0.325 mAP@0.5 at 49.5 FPS [arxiv:2205.14942 tab. 5]. Post-processing parameters are not stated [arxiv:2205.14942 alg. 2] |
| LiDAR only, sensor-native range image, and a spread for each box is wanted | LaserNet | 85.34 AP0.7 on the ATG4D validation set, against 80.99 for PIXOR [arxiv:1903.08701 tab. 1]. On KITTI it trails BEV methods, 73.77 against 77.05 for PIXOR [arxiv:1903.08701 tab. 4]. The dataset's access route is not stated [arxiv:1903.08701 §4.1] |
| Crowd counts where detection and density each fail in part of the range | DecideNet, with a per-pixel blend of the two maps | 1.52 MAE on Mall, against 1.82 for Exemplary Density [arxiv:1712.06679 tab. 1]. CP-CNN reports lower MAE on ShanghaiTech PartB, 20.1 against 21.53 [arxiv:1712.06679 tab. 2] |

## Implementation sketch

The shared core of a 2D detector, with the points where the variants differ.
Same style as paper cards: Python-flavoured, lines of at most 80 characters.

```pseudocode
# Common core of 2D object detection  [arxiv:1805.04953 §3] [arxiv:2211.12860 §3.1]
# Input:  image (H, W, 3); model with backbone, head and post-processing
# Output: list of (class, box, score)

def detect(image, model, score_threshold):
    # Shared convolutional features  [arxiv:1805.04953 §3]
    feats = model.backbone(image)
    # Candidates: anchors (SSD, YOLO), proposals (Faster R-CNN) or
    # learned queries (DETR-style)  [arxiv:1802.09567 §4.1] [arxiv:2211.12860 §3.1]
    candidates = model.make_candidates(feats)
    # Each candidate gets class scores and box offsets  [arxiv:1805.04953 §3.3]
    scores, boxes = model.head(feats, candidates)
    # Keep confident candidates  [arxiv:1802.09567 §4.1]
    keep = scores.max(axis=-1) > score_threshold
    scores, boxes = scores[keep], boxes[keep]
    # Dense and two-stage detectors drop duplicates with NMS, IoU 0.2 in RGB-N
    # [arxiv:1805.04953 §3.4]. Base DETR-style models skip this step
    # [arxiv:2211.12860 §1]
    if model.needs_nms:
        boxes, scores = nms(boxes, scores, iou_threshold=0.2)
    return boxes, scores
```

## Contradictions

- **No shared benchmark.** The cards use different datasets and metrics:
  per-class recall, precision and an undefined "accuracy" on road damage
  [arxiv:1801.09454 §5], plate read rate on frames [arxiv:1802.09567 §5], pixel
  F1 on forgery benchmarks [arxiv:1805.04953 §4.2], COCO AP for DETR-style
  models [arxiv:2211.12860 §4.2] and [arxiv:2011.10881 tab. 1], AP on AI-TOD
  [arxiv:2208.08738 tab. 5], ScanNet and SUN RGB-D mAP [arxiv:2104.00678 tab. 1],
  mAP on nuScenes [arxiv:2205.13790 tab. 1], Wilderness Impact and A-OSE for open-world detection [arxiv:2103.02603 §5.1], whole-scene mAP at IoU 0.25 on
  ARKitScenes [arxiv:2111.08897 tab. 2], and AP on simulated LiDAR
  [arxiv:2109.07644 §5.2], and for 3D, KITTI at 11 or 40 recall positions, nuScenes
  NDS and mAP, and Waymo APH [arxiv:2104.10330 §4.2] [arxiv:2110.06922 tab. 1]
  [arxiv:2203.17270 tab. 3]. Their numbers are not comparable across papers.
- **Why DETR converges slowly.** TSP attributes the slow training mainly to the
  cross-attention module [arxiv:2011.10881 §1]. H-DETR attributes it to few
  positive queries per object under one-to-one matching [arxiv:2207.13080 §1].
  Both fixes help, but the cards do not test each cause on its own.
- **Road-damage headline.** The authors say SSD MobileNet is better overall and
  that the best class exceeds 0.75 recall and precision at 1.5 s on a phone
  [arxiv:1801.09454 §5] [arxiv:1801.09454 §6]. The table shows Inception V2 is
  better on several classes, and only D43 with Inception V2 clears 0.75 on both
  [arxiv:1801.09454 tab. 2].
- **Best polyp detector.** The authors name YOLOv4 as the best detector
  [arxiv:2011.07631 §VI]. Their table gives RetinaNet with ResNet-101 the highest
  AP, 0.8745 against 0.8513 for YOLOv4 [arxiv:2011.07631 tab. III]. The authors
  weigh speed as well as AP.
- **ColonSegNet detection.** The card reports AP 0.8000 at 180 FPS for ColonSegNet
  in detection, but the public repository implements only segmentation
  [arxiv:2011.07631 tab. III] [code: README.md#L23-L25]. The detection numbers
  cannot be reproduced from the release.
- **NMS-free claim.** The base Co-DETR method needs no NMS
  [arxiv:2211.12860 §1]. Its strongest Co-DINO results do use NMS
  [arxiv:2211.12860 §4.3], so gains in that comparison mix the training scheme
  with post-processing.
- **Which YOLO version is best.** [YOLO Evolution](/papers/2411.00201) puts YOLO11
  and YOLOv12 level on accuracy and YOLO11 first on balance. [YOLO11 and YOLOv8 on
  orchard fruit](/papers/2410.19869) finds YOLO11m-seg most accurate and
  YOLOv8n-seg fastest. The datasets, tasks and hardware differ, and each card
  reports one run per model [arxiv:2411.00201 tab. 4] [arxiv:2410.19869 tab. 2]
  [arxiv:2410.19869 tab. 3].
- **Text and figures in the YOLO benchmark.** YOLO Evolution gives YOLOv5ul's
  traffic-sign mAP50 as 0.866 in the text and 0.886 in its figure [arxiv:2411.00201
  fig. 12]. Its text says the tables use the test split, while the authors' code
  reports validation metrics, so the split behind the numbers is unclear
  [arxiv:2411.00201 §3.1].
- **Small-object gains.** YOLO-Z's abstract gives up to 6.9% relative gain on small
  objects, while its Table 2 gives 5.46% to 7.73% [arxiv:2112.11798 tab. 2]. Only the
  table is used on this page.
- **ReDet's abstract against its tables.** The abstract gives gains of 1.2, 3.5
  and 2.6 mAP over the previous best. Against the strongest other row in its
  tables the gains are 0.68, 1.90 and 2.62, and on HRSC2016 VOC2007 only 0.29
  [arxiv:2103.07733 §1] [arxiv:2103.07733 tab. 6] [arxiv:2103.07733 tab. 8]. The
  abstract's model sizes, 313 Mb and 121 Mb, appear in no table row
  [arxiv:2103.07733 §1].
- **Dynamic Head's cost.** The abstract says the head adds no computational
  overhead. Table 2 shows +21.50 GFLOPs at six blocks, and the text calls this
  negligible [arxiv:2106.08322 tab. 2] [arxiv:2106.08322 §4.2]. Its 54.0 AP
  result uses test-time augmentation and a DCN backbone, so the fairer
  comparison with HTC++ is its 58.7 AP without extra data [arxiv:2106.08322
  §4.3] [arxiv:2106.08322 tab. 7].
- **Summary table against the text, YOLO review.** The review's summary table
  and its prose give different AP for several versions, and some table entries
  come from other settings. The table gives 33.0 for YOLOv3, where the text gives
  36.2 for YOLOv3-spp [arxiv:2304.00501 tab. 4] [arxiv:2304.00501 §6.3]. The
  table's 55.8 for YOLOv5 is the 1536-pixel figure with test-time augmentation,
  while the 640-pixel figure in the text is 50.7 [arxiv:2304.00501 §9]. Use the
  original papers for exact values.
- **Schedules in the backbone comparison.** The PVT v2 text says its comparisons
  use 1x or 3x schedules, but its Table 4 does not say which one was used for
  ATSS, GFL or Cascade Mask R-CNN [arxiv:2106.13797 §4.2] [arxiv:2106.13797
  tab. 4]. Compare those rows only with that caveat.
- **PVT v2 learning rate.** The text gives 1e-4 for Mask R-CNN at 1x, while the
  released 1x config sets 2e-4 [arxiv:2106.13797 §4.2] [code:
  detection/configs/mask_rcnn_pvt_v2_b2_fpn_1x_coco.py#L20]. Pick one and state
  which.
- **RFAConv's replacement scope.** The paper says every 3×3 convolution in the
  YOLOv5 and YOLOv8 yaml files is replaced [arxiv:2304.03198 §4.3]. The shipped
  yaml files replace only some of them [code: Detection/YOLOv5/models/yolov5n.yaml
  (NovelConv)#L16-L22] [code: Detection/YOLOv8n.yaml#L13-L19], and the NovelConv
  class they use is not in the repository. The detection numbers therefore cannot
  be reproduced from that code [arxiv:2304.03198 §4.3].
- **DETR3D's depth claim.** The authors say DETR3D's mATE is worse than
  FCOS3D's, and attribute the gap to FCOS3D's depth supervision
  [arxiv:2110.06922 §4.2]. The card records this as the authors' suspicion, not
  a measured cause.
- **Few-shot headline.** The text says FSCE gains up to 8.8 nAP50 on split 3 at
  10 shots. That gain is measured against MPSR, at 49.7. TFA w/ fc is 50.2 in
  the same column, so the margin to the best unstarred row is 8.3 [arxiv:2103.05950
  §4.2] [arxiv:2103.05950 tab. 2].
- **COD10K test size.** The COD-Rank card gives COD10K 2,026 test images
  [arxiv:2103.04011 §1] [arxiv:2103.04011 §2]. SINet-V2 gives 4,000 test images in
  its split [arxiv:2102.10274 §3.3] [arxiv:2102.10274 tab. 1], and 2,026 in its
  results [arxiv:2102.10274 §5.2]. Compare COD10K numbers only with the split
  stated.
- **MARTENS failure counts.** The results tables give 50% (use case 1) and 26%
  (use case 2) more failures than random search [arxiv:2410.19277 tab. 2]
  [arxiv:2410.19277 tab. 3]. The paper's abstract and conclusion state the gain as
  a range (25% to 50%, or 26% to 50%), and the printed row sums do not match the
  printed totals [arxiv:2410.19277 tab. 2]. Both figures are shown; they are not
  reconciled here.
- **Edge YOLO's model size.** The abstract gives 25.67 MB, while Table 5 gives
  25.27 MB [arxiv:2205.14942 tab. 5]. The table value is used on this page.
- **Edge YOLO's KITTI mAP.** The text gives KITTI mAP values of 0.7904 and 0.8473
  [arxiv:2205.14942 §IV-C], while Table 6 lists YOLOv3-Tiny at 0.7804
  [arxiv:2205.14942 tab. 6]. The table rows are used on this page.
- **Survey text against its own table.** The 3D survey gives PV-RCNN++ an easy
  AP3D of 90.90 in its text, while its table gives PV-RCNN++ 90.14 and gives
  90.90 to Voxel R-CNN [arxiv:2206.09474 §10.1.1] [arxiv:2206.09474 tab. 16].

## Papers

- [Road Damage Detection Using Deep Neural Networks with Images Captured Through a Smartphone](/papers/1801.09454): SSD baselines on a public road-damage dataset, with phone and GPU timings.
- [FSSD: Feature Fusion Single Shot Multibox Detector](/papers/1712.00960): fuses three SSD source layers into one map before the detection pyramid.
- [xView: Objects in Context in Overhead Imagery](/papers/1802.07856): an overhead detection dataset with an SSD baseline and three training-set variants.
- [Real-time ALPR with YOLO (Laroca et al.)](/papers/1802.09567): a cascade of YOLO-style detectors for plates, with per-stage timing.
- [Real-Time Polyp Detection, Localization and Segmentation in Colonoscopy Using Deep Learning](/papers/2011.07631): benchmark of detectors and segmenters on a colonoscopy dataset, with a fast proposed network.
- [Rethinking Transformer-based Set Prediction for Object Detection](/papers/2011.10881): a transformer encoder over selected features, trained with a restricted matching loss.
- [End-to-end Interpretable Neural Motion Planner](/papers/2101.06679): one network for LiDAR-based detection, motion and a planning cost map.
- [Unbiased Teacher for Semi-Supervised Object Detection](/papers/2102.09480): teacher-student pseudo-labelling for Faster R-CNN with few labels.
- [ICDAR2019 Competition on Scanned Receipt OCR and Information Extraction](/papers/2103.10213): text localisation, OCR and field extraction on receipts, with the top entries.
- [Group-Free 3D Object Detection via Transformers](/papers/2104.00678): transformer decoder over all points for 3D indoor detection, with no point grouping.
- [A Comprehensive Survey of Scene Graphs: Generation and Application](/papers/2104.01111): surveys relation detection on top of object detection, with reported benchmark numbers.
- [SeaDronesSee: A Maritime Benchmark for Detecting Humans in Open Water](/papers/2105.01922): drone dataset for detecting and tracking people and boats in open water.
- [OPV2V](/papers/2109.07644): a simulated benchmark and intermediate-fusion pipeline for 3D detection between cars.
- [Autonomous Vehicles on the Edge: A Survey on Autonomous Vehicle Racing](/papers/2202.07008): survey of perception, planning and control for racing cars, with open problems.
- [Knowledge Distillation from A Stronger Teacher](/papers/2205.10536): Pearson-correlation distillation, with COCO detection results.
- [BEVFusion: A Simple and Robust LiDAR-Camera Fusion Framework](/papers/2205.13790): camera and LiDAR streams in a shared BEV grid, robust to LiDAR loss.
- [DETRs with Hybrid Matching](/papers/2207.13080): a training-only one-to-many group added to DETR-style detectors.
- [RFLA: Gaussian Receptive Field based Label Assignment for Tiny Object Detection](/papers/2208.08738): label assignment by receptive-field distance, for tiny objects.
- [Transformers in Remote Sensing: A Survey](/papers/2209.01206): survey of transformer methods for remote sensing, with reported detection numbers.
- [Florence-2: Advancing a Unified Representation for a Variety of Vision Tasks](/papers/2311.06242): one model for captioning, detection and grounding, written as text.
- [DETRs with Collaborative Hybrid Assignments Training](/papers/2211.12860): auxiliary one-to-many heads that speed up DETR-style training, with the inference path unchanged.
- [YOLO-LITE: A Real-Time Object Detection Algorithm Optimized for Non-GPU Computers](/papers/1811.05588): a small detector found from Tiny-YOLOv2, for laptops and phones without a GPU.
- [microYOLO: Towards Single-Shot Object Detection on Microcontrollers](/papers/2408.15865): YOLO-style detector for a Cortex-M7 board, with pruning and 8-bit quantization.
- [ARKitScenes: A Diverse Real-World Dataset For 3D Indoor Scene Understanding Using Mobile RGB-D Data](/papers/2111.08897): indoor RGB-D captures with 3D furniture boxes and three detector baselines.
- [Anomaly Detection in Autonomous Driving: A Survey](/papers/2204.07974): survey of corner-case detection across camera, lidar, radar and object-level data.
- [Metrics reloaded: Recommendations for image analysis validation](/papers/2206.01653): expert-consensus decision guide for choosing validation metrics, including detection.
- [RDD2022: A multi-national image dataset for automatic Road Damage Detection](/papers/2209.08538): six-country road-damage dataset with four classes and no detector baseline.
- [OrientedFormer](/papers/2409.19648): DETR-style detector for rotated boxes in aerial images.
- [Visual-RFT: Visual Reinforcement Fine-Tuning](/papers/2503.01785): reinforcement fine-tuning of a vision-language model with a box reward.
- [Synscapes: A Photorealistic Synthetic Dataset for Street Scene Parsing](/papers/1810.08705): synthetic street scenes with instance, depth and box labels.
- [Monitoring COVID-19 social distancing with person detection and tracking via fine-tuned YOLO v3 and Deepsort techniques](/papers/2005.01385): YOLO v3 with DeepSORT for a distance monitor, with unvalidated grouping.
- [Equalization Loss v2: A New Gradient Balance Approach for Long-tailed Object Detection](/papers/2012.08548): per-class gradient reweighting for sigmoid detectors, in one stage.
- [YOLO5Face: Why Reinventing a Face Detector](/papers/2105.12931): YOLOv5 adapted to faces, with a landmark head and small mobile models.
- [Disentangle Your Dense Object Detector](/papers/2107.02963): separate assignment, deformable head layers and level weights for one-stage detectors.
- [PubTables-1M: Towards comprehensive table extraction from unstructured documents](/papers/2110.00061): table dataset with DETR baselines for table structure.
- [YOLO-Z: Improving small object detection in YOLOv5 for autonomous vehicles](/papers/2112.11798): YOLOv5 changes for small traffic cones, with a higher-resolution map.
- [PolarFormer: Multi-camera 3D Object Detection with Polar Transformer](/papers/2206.15398): camera-only 3D detection on a polar bird's-eye-view grid.
- [DeepInteraction: 3D Object Detection via Modality Interaction](/papers/2208.11112): LiDAR and camera features kept apart and exchanged during detection.
- [SportsMOT: A Large Multi-Object Tracking Dataset in Multiple Sports Scenes](/papers/2304.05170): tracking dataset for players in three sports, with a tracker that adds appearance.
- [What is YOLOv8: An In-Depth Exploration of the Internal Features of the Next-Generation Object Detector](/papers/2408.15857): descriptive review of YOLOv8, with no experiments of its own.
- [Comparing YOLOv11 and YOLOv8 for instance segmentation of occluded and non-occluded immature green fruits in complex orchard environment](/papers/2410.19869): YOLO11-seg and YOLOv8-seg benchmark on orchard apples.
- [YOLO Evolution: A Comprehensive Benchmark and Architectural Review of YOLOv12, YOLO11, and Their Previous Versions](/papers/2411.00201): 33 YOLO checkpoints on three datasets, one training setting.
- [Joint 3D Proposal Generation and Object Detection from View Aggregation](/papers/1712.02294): LiDAR and RGB fused into 3D car boxes through a proposal network.
- [Deep learning in agriculture: A survey](/papers/1807.11809): survey of 40 agricultural deep-learning studies, with tabulated settings and results.
- [ReDet: A Rotation-equivariant Detector for Aerial Object Detection](/papers/2103.07733): rotation-equivariant backbone and RiRoI Align for oriented aerial boxes.
- [Swin Transformer: Hierarchical Vision Transformer using Shifted Windows](/papers/2103.14030): windowed transformer backbone, used with detectors on COCO.
- [BADet: Boundary-Aware 3D Object Detection from Point Clouds](/papers/2104.10330): graph refinement of LiDAR proposals for 3D detection.
- [Dynamic Head: Unifying Object Detection Heads with Attentions](/papers/2106.08322): stacked scale, spatial and task attention blocks as a detector head.
- [3D Object Detection for Autonomous Driving: A Survey](/papers/2106.10823): survey of 3D detection by input type, with a case study of fifteen detectors.
- [DETR3D: 3D Object Detection from Multi-view Images via 3D-to-2D Queries](/papers/2110.06922): camera-only 3D detection from learned object queries, with no NMS.
- [BEVFormer: Learning Bird's-Eye-View Representation from Multi-Camera Images via Spatiotemporal Transformers](/papers/2203.17270): BEV grid from surround cameras, with spatial and temporal attention.
- [A Comprehensive Review of YOLO Architectures in Computer Vision: From YOLOv1 to YOLOv8 and YOLO-NAS](/papers/2304.00501): review of YOLO versions with a summary table of reported AP.
- [YOLOv10: Real-Time End-to-End Object Detection](/papers/2405.14458): YOLO without NMS, with one-to-many and one-to-one heads.
- [YOLOv11: An Overview of the Key Architectural Enhancements](/papers/2410.17725): architectural overview of YOLOv11, with no experiments of its own.
- [A Survey of Modern Deep Learning based Object Detection Models](/papers/2104.11892): survey of detector families, datasets and metrics, with compiled author-reported numbers.
- [Conformer: Local Features Coupling Global Representations for Visual Recognition](/papers/2105.03889): hybrid CNN and transformer backbone, with COCO detection results.
- [GenEval: An Object-Focused Framework for Evaluating Text-to-Image Alignment](/papers/2310.11513): text-to-image benchmark whose checks rely on a COCO-trained detector.
- [Categorical Depth Distribution Network for Monocular 3D Object Detection](/papers/2103.01100): camera-only car detector that lifts per-pixel depth distributions into bird's-eye view.
- [Voxel Transformer for 3D Object Detection](/papers/2109.02497): Transformer backbone over non-empty voxels for LiDAR detection, with a hash-table neighbour search.
- [TopoLogic: An Interpretable Pipeline for Lane Topology Reasoning on Driving Scenes](/papers/2405.14747): lane-to-lane connection scores from endpoint gaps and query similarity, fused and passed to a graph layer.
- [Learning Rich Features for Image Manipulation Detection](/papers/1805.04953): Faster R-CNN with RGB and noise streams for forgery localisation.
- [PVT v2: Improved Baselines with Pyramid Vision Transformer](/papers/2106.13797): pyramid transformer backbone with pooled-key attention, overlapping patches and a depthwise convolution in the feed-forward block.
- [RFAConv: Receptive-Field Attention Convolution for Improving Convolutional Neural Networks](/papers/2304.03198): convolution with per-window attention over each receptive field, as a drop-in replacement.
- [Towards Open World Object Detection](/papers/2103.02603): Faster R-CNN variant that flags unknown objects and learns them incrementally.
- [RTMDet: An Empirical Study of Designing Real-Time Object Detectors](/papers/2212.07784): one-stage detector family from tiny to extra-large, with a 5 by 5 depthwise block and soft label assignment.
- [VisionLLM: Large Language Model is also an Open-Ended Decoder for Vision-Centric Tasks](/papers/2305.11175): detection and other vision tasks written as text and decoded by an LLM.
- [POI: Multiple Object Tracking with High Performance Detection and Appearance Feature](/papers/1610.06136): Faster R-CNN detections with a learned appearance feature for tracking.
- [Video Salient Object Detection via Fully Convolutional Networks](/papers/1702.00871): static and dynamic FCNs on frame pairs, trained on warped still images.
- [Deep Visual Attention Prediction](/papers/1705.02544): eye-fixation model with three deconvolution side outputs and a learned fusion.
- [Structure-measure: A New Way to Evaluate Foreground Maps](/papers/1708.00786): region and object structural similarity score for non-binary maps.
- [Disentangling Monocular 3D Object Detection](/papers/1905.12365): monocular car detector with a split 3D loss and a 40-point AP variant.
- [Rethinking RGB-D Salient Object Detection: Models, Data Sets, and Large-Scale Benchmarks](/papers/1907.06781): three-stream RGB-D model with a test-time depth gate and a phone dataset.
- [EGNet: Edge Guidance Network for Salient Object Detection](/papers/1908.08297): salient-edge path that guides each object scale of a salient-object network.
- [Concealed Object Detection](/papers/2102.10274): SINet-V2 for camouflaged objects, with the COD10K dataset.
- [Simultaneously Localize, Segment and Rank the Camouflaged Objects](/papers/2103.04011): fixation, camouflage mask and difficulty rank in one network; ranking code absent.
- [FSCE: Few-Shot Object Detection via Contrastive Proposal Encoding](/papers/2103.05950): contrastive head on RoI proposals for few-shot fine-tuning.
- [Anabranch Network for Camouflaged Object Segmentation](/papers/2105.09451): classifier stream that scales the map of a saliency segmenter; CAMO dataset.
- [YOLOP: You Only Look Once for Panoptic Driving Perception](/papers/2108.11250): one encoder with heads for vehicles, drivable area and lane lines.
- [YOLO-Pose: Enhancing YOLO for Multi Person Pose Estimation Using Object Keypoint Similarity Loss](/papers/2204.06806): YOLO head with keypoints per anchor and an OKS-based keypoint loss.
- [MARTENS](/papers/2410.19277): genetic search for failing arm scenes in simulation, then detector repair and real-image fine-tuning.
- [Moving Forward](/papers/2411.10291): survey and benchmark of autonomous-driving software and hardware, with a detection results table.
- [Pseudo-LiDAR++: Accurate Depth for 3D Object Detection in Autonomous Driving](/papers/1906.06310): stereo depth trained on depth, corrected with sparse LiDAR points by graph propagation.
- [CenterFusion: Center-based Radar and Camera Fusion for 3D Object Detection](/papers/2011.04841): radar points matched to camera detections and painted as depth and velocity maps.
- [Inner-IoU](/papers/2311.02877): box regression loss computed on scaled auxiliary boxes, with no added term.
- [DecideNet (crowd counting)](/papers/1712.06679): detector boxes as Gaussian bumps, blended with a regression map by a per-pixel attention net.
- [LaserNet: An Efficient Probabilistic 3D Object Detector for Autonomous Driving](/papers/1903.08701): range-image LiDAR detector with a mixture of box distributions and variance-aware NMS.
- [A Review on Deep Learning in UAV Remote Sensing](/papers/2101.10861): literature review of deep learning on drone imagery, tallied by network, task and sensor.
- [A Survey of Deep Learning Techniques for Weed Detection from Images](/papers/2103.01415): survey of 70 deep-learning weed-detection papers and 19 public datasets.
- [Edge YOLO: Real-Time Intelligent Object Detection System Based on Edge-Cloud Cooperation in Autonomous Vehicles](/papers/2205.14942): slimmed YOLOv4 for an in-vehicle Jetson board, retrained in the cloud.
- [3D Object Detection for Autonomous Driving: A Comprehensive Survey](/papers/2206.09474): survey of 3D detection by sensor, with compiled author-reported scores.
- [YOLO advances to its genesis: a decadal and comprehensive review of the You Only Look Once (YOLO) series](/papers/2406.19407): review of YOLOv1 to YOLOv12 and variant detectors, with no experiments
- [Learning Salient Boundary Feature for Anchor-free Temporal Action Localization](/papers/2103.13137): anchor-free temporal action localiser with boundary pooling.
- [Image Data Augmentation for Deep Learning: A Survey](/papers/2204.08610): augmentation survey with its own detection, segmentation and classification runs.
- [V2V4Real: A Real-world Large-scale Dataset for Vehicle-to-Vehicle Cooperative Perception](/papers/2303.07601): two-car LiDAR dataset with cooperative detection, tracking and Sim2Real benchmarks. of its own.