---
kind: concept
slug: salient-object-detection
title: Salient object detection
summary: >-
  Salient object detection predicts a per-pixel map of the most eye-catching object, scored against a binary mask. S-measure scores structure where AP, AUC and Fbw only count pixels. EGNet adds an edge branch (max F 0.893 on DUTS-TE). Variants add depth (D3Net, SwinNet), thermal (SwinNet), camouflage (SINet-V2, ANet, COD-Rank), eye fixations (DVA) and video (no optical flow).
tags: [saliency, segmentation, evaluation, computer-vision, camouflage, rgb-d, rgb-t, transformer, eye-fixation, video]
papers: ["1702.00871", "1705.02544", "1708.00786", "1907.06781", "1908.08297", "2102.10274", "2103.04011", "2105.09451", "2204.05585"]
related: [segmentation-evaluation, structural-similarity, multi-scale-feature-fusion, class-balanced-loss, multi-task-learning, multimodal-fusion, deep-supervision, optical-flow, channel-attention, hierarchical-vision-transformer]
status: draft
updated_at: 2026-11-01
---

## Overview

Salient object detection (SOD) takes an image and outputs a map that gives
every pixel a value for how likely it is to be part of the main foreground
object. Models produce non-binary maps, while the ground truth (GT) is
normally a binary mask [arxiv:1708.00786 §1]. This makes evaluation a design
problem of its own: how should a soft map be scored against a hard mask?

The wiki has nine sources so far. One concerns the evaluation question, and
the rest are models for image SOD or for neighbouring settings. The
Structure-measure paper argues
that the common scores (AP, AUC and the weighted F-measure Fbw) reduce a map
to per-pixel TP/TN/FP/FN counts and so cannot tell apart errors that keep an
object's shape from errors that break it [arxiv:1708.00786 §2, §3]. Its
example: a square hole inside a digit and one cutting through the digit's
outline give identical confusion counts [arxiv:1708.00786 fig. 3]. It
proposes S-measure, which scores region-level and object-level structure
instead [arxiv:1708.00786 §4].

The EGNet paper is about a model. It trains one network to predict the object
and its outline together, and feeds the outline features back into every
scale of the object prediction [arxiv:1908.08297 §3.2, §3.3]. Its results are
reported with max F-measure, MAE and S-measure.

The other sources extend the same per-pixel prediction to neighbouring
settings. Two add a depth map to the colour image: D3Net [arxiv:1907.06781 §4]
and SwinNet, which also takes thermal images [arxiv:2204.05585 §3.1]. Human eye
fixations [arxiv:1705.02544 §1] and a neighbouring video frame, read alongside
a static saliency map [arxiv:1702.00871 §3.1], are two more. The remaining three
apply it to camouflaged objects, which share texture and
colour with their background, so that low-level edge and saliency cues mislead
detectors [arxiv:2102.10274 §1, §2] [arxiv:2105.09451 §1, §2.2]
[arxiv:2103.04011 §1]. SINet-V2 is a search-and-identification cascade for
concealed objects. The Anabranch Network (ANet) multiplies a saliency-style
map by an image-level camouflage probability, so the output is scaled by how
likely camouflage is present at all [arxiv:2105.09451 §4.2]. COD-Rank adds a
third head that ranks how hard each camouflaged object is to spot
[arxiv:2103.04011 §3.3].

The mask-based measures below share one contract: a non-binary prediction is
compared with a binary GT of the same size, and the output is a scalar per
image that is averaged over a dataset. Eye-fixation work is the exception. Its
GT is a fixation map and its metrics (NSS, CC, AUC variants, EMD) are different
[arxiv:1705.02544 §4.2], so its numbers do not sit with the mask-based ones.

The S-measure paper ranks ten published models (ST, DRFI, DSR, DCL, rfcn, MC,
MDF, DISC, DHS, ELD) with its measure [arxiv:1708.00786 §5.5]. Their methods
are outside this wiki and have no cards.

## Variants

### Pixel-wise measures: AP, AUC, Fbw

These are the baselines the S-measure paper compares against
[arxiv:1708.00786 §5, tab. 1]. Each treats pixels independently: the map is
compared to the GT through confusion counts, either over thresholds (AP,
AUC) or with per-pixel weighting (Fbw) [arxiv:1708.00786 §2]. The card does
not give their formulas. The paper's criticism is that structure does not
enter the score, and it shows Fbw ranking a blurry detection that only
covers the top of a pyramid above one that keeps the whole shape
[arxiv:1708.00786 §3, fig. 3].

### Structure-measure (S-measure)

`S = α·S_o + (1 − α)·S_r` with `α = 0.5` [arxiv:1708.00786 eq. 10, §4.3].

- **Region term `S_r`:** split map and GT into four blocks at the centroid
  of the GT foreground, compute SSIM on each block without the usual
  stabilising constants, and sum with block weights
  [arxiv:1708.00786 §4.1, eq. 3, eq. 4].
- **Object term `S_o`:** inside the GT foreground, reward a high and evenly
  spread predicted mean; on the GT background, do the same on the inverted
  map; weight the two by the foreground area fraction
  [arxiv:1708.00786 §4.2, eq. 7, eq. 8, eq. 9].

The paper's experiments use `λ = 0.5` and `K = 4` blocks
[arxiv:1708.00786 §5]. Runtime is 5.3 ms per image on one CPU thread in
Matlab [arxiv:1708.00786 §5]. Card:
[Structure-measure: A New Way to Evaluate Foreground Maps](/papers/1708.00786).

Reported comparison across four datasets (lower is better on all three
meta-measures; MM1 = 1 − Spearman ρ against a SalCut application ranking,
MM2 = % of images where a generic map beats the state-of-the-art mean,
MM3 = % of cases where a wrong GT raises the score)
[arxiv:1708.00786 §5.1, §5.2, §5.3, tab. 1]:

| Measure | PASCAL-S MM1 | MM2 (%) | MM3 (%) | ECSSD MM1 | MM2 (%) | MM3 (%) | SOD MM1 | MM2 (%) | MM3 (%) | HKU-IS MM1 | MM2 (%) | MM3 (%) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| AP | 0.452 | 12.1 | 5.50 | 0.449 | 9.70 | 3.32 | 0.504 | 9.67 | 7.69 | 0.518 | 3.76 | 1.25 |
| AUC | 0.449 | 15.8 | 8.21 | 0.436 | 12.1 | 4.18 | 0.547 | 14.0 | 8.27 | 0.519 | 7.02 | 2.12 |
| Fbw | 0.365 | 7.06 | 1.05 | 0.401 | **3.00** | 0.84 | 0.384 | 16.3 | 0.73 | 0.498 | 0.36 | 0.26 |
| S-measure | **0.320** | **4.59** | **0.34** | **0.312** | 3.30 | **0.47** | **0.349** | **9.67** | **0.60** | **0.424** | **0.34** | **0.08** |

[arxiv:1708.00786 tab. 1]

In a user study with 45 viewers on 50 image pairs, people preferred the map
that S-measure ranked first in 63.69% of trials against Fbw, 72.11% against
AP and 73.56% against AUC [arxiv:1708.00786 §5.6].

### Edge-guided network: EGNet

EGNet runs two paths over one backbone [arxiv:1908.08297 §3.2]:

- **Object path:** a top-down pass over side features C3 to C6, where each
  level adds the upsampled level above it before a three-layer enhancement
  block [arxiv:1908.08297 §3.2.1, eq. 3, tab. 1].
- **Edge path:** starts from the C2 side features, adds the top-level
  location, and is supervised by a cross-entropy edge loss
  [arxiv:1908.08297 §3.2.2, eq. 2, eq. 4].

The edge features are added to each object-scale feature. Each of the four
guided scales then predicts its own map, and the four maps are fused into
the output [arxiv:1908.08297 §3.3, eq. 7, eq. 9]. Training sums the edge
loss, deep supervision on every object side output and a loss on each guided
map [arxiv:1908.08297 eq. 6, eq. 10]. At test time only the fused map is
used [arxiv:1908.08297 §4.1]. The paper says its results come without any
pre- or post-processing, unlike several baselines marked in its table
[arxiv:1908.08297 §4.4, tab. 2].

Author-reported max F-measure (β² = 0.3) [arxiv:1908.08297 tab. 2]:

| Backbone | ECSSD | PASCAL-S | DUT-O | HKU-IS | SOD | DUTS-TE |
|---|---|---|---|---|---|---|
| ResNet | 0.943 | 0.869 | 0.842 | 0.937 | 0.890 | 0.893 |
| VGG | 0.941 | 0.863 | 0.826 | 0.929 | 0.869 | 0.880 |

The strongest baseline on DUTS-TE is PiCANet with ResNet, at 0.863 with
post-processing [arxiv:1908.08297 tab. 2]. The ablation trains on DUTS-TR
and tests on SOD and DUTS-TE. Adding the top-down edge path to the U-Net
baseline lifts DUTS-TE max F from 0.855 to 0.879, and the full model reaches
0.893. Swapping in NLDF's edge penalty gives 0.866 [arxiv:1908.08297 §4.3,
tab. 3]. The paper also reports the edge maps' own max F on DUTS, 0.539
against 0.429 for NLDF's edges [arxiv:1908.08297 §4.3.3, tab. 4]. EGNet also
reports S-measure with γ = 0.5 [arxiv:1908.08297 §4.2, eq. 13]; its ResNet
DUTS-TE value is 0.875 [arxiv:1908.08297 tab. 2]. Card:
[EGNet: Edge Guidance Network for Salient Object Detection](/papers/1908.08297).

### RGB-D salient object detection: D3Net

D3Net adds a depth map to the colour image [arxiv:1907.06781 §4]. It has three
VGG-16 feature streams: one on RGB, one on RGB-D stacked as four channels, and
one on depth alone [arxiv:1907.06781 §4.1]. Each stream builds a feature
pyramid and predicts its own map. At test time a depth depurator unit (DDU)
compares the RGB-D map with the depth-only map by mean absolute difference.
It keeps the RGB-D map when the difference is at most t = 0.15, and otherwise
keeps the RGB map [arxiv:1907.06781 §4.2, eq. 2, eq. 3, §4.3]. The premise is
that a poor depth map gives a cluttered scene, while a good one gives a clear
object with closed boundaries [arxiv:1907.06781 §4.2, fig. 7].

Each stream is trained alone with pixel-wise binary cross-entropy on the
NJU2K and NLPR training pairs (1,485 and 700 pairs) [arxiv:1907.06781 §4.3,
eq. 4]. The abstract says the parts are learned jointly, but the released code
trains each stream in its own run (see the card's Paper vs code section)
[arxiv:1907.06781 abstract].

Author-reported S-measure on the test sets, D3Net against CPFP, the best
published method in the paper's comparison [arxiv:1907.06781 §5.5, tab. 4]:
NJU2K 0.900 against 0.879, NLPR 0.912 against 0.888, SIP 0.860 against 0.850.
On LFSD, CPFP is ahead: 0.828 against 0.825. The abstract's claim of exceeding
all prior work on all five metrics is therefore not borne out on every row
[arxiv:1907.06781 abstract, tab. 4]. The ablation shows the gate is not
always a gain: on SIP, D3Net scores 0.860, below the RGB-D stream alone at
0.862 [arxiv:1907.06781 tab. 5]. Card:
[Rethinking RGB-D Salient Object Detection: Models, Data Sets, and Large-Scale Benchmarks](/papers/1907.06781).

### Transformer RGB-D and RGB-T: SwinNet

SwinNet runs two Swin-B encoders, one on the colour image and one on the second
modality [arxiv:2204.05585 §3.1, §3.2]. The paper describes the RGB-D case and
says RGB-T uses the same design [arxiv:2204.05585 §3.1]. Swin's shifted windows
let neighbouring windows exchange information, which the authors set against
the limited long-range context of CNN backbones [arxiv:2204.05585 §1].

At each of the four levels, a spatial attention map computed from the product
of the two streams aligns them. Each stream is then re-weighted by channel
attention that is computed on the aligned features and applied to the unaligned
originals [arxiv:2204.05585 eq. 1 to eq. 6]. An edge feature built from the
three shallowest levels of the depth stream guides the decoder
[arxiv:2204.05585 §3.4, eq. 7, eq. 8]. Each level fuses the two streams by
concatenating their sum and their element-wise product [arxiv:2204.05585 eq. 9].
Training sums a binary cross-entropy loss on the edge map and one on the
saliency map. The edge targets are Canny edges of the ground truth
[arxiv:2204.05585 eq. 12, eq. 14, eq. 16, §3.6.1].

Training uses 384 × 384 inputs, Adam with learning rate 5 × 10⁻⁵ and batch size
3. The authors report about 26 hours on one RTX 2080 Ti [arxiv:2204.05585 §4.2].

Author-reported S-measure is 0.941 on NLPR and 0.935 on NJU2K, against 0.931
and 0.922 for VST, the transformer baseline the paper singles out
[arxiv:2204.05585 §4.3, tab. 1]. On RGB-T, S-measure is 0.912 on VT5000 against
0.883 for CGFNet [arxiv:2204.05585 tab. 2]. The RGB-T text says every metric is
best, but CGFNet has the higher F-measure on VT1000 (0.906 against 0.896)
[arxiv:2204.05585 §4.3.2, tab. 2]. In the paper's backbone ablation, Swin-B has
the highest S-measure on all four RGB-D test sets against six other backbones
[arxiv:2204.05585 tab. 3].

Depth does not always help. On STERE, RGB alone scores S 0.923 against 0.919
for RGB plus depth, which the paper puts down to low-quality depth maps
[arxiv:2204.05585 §4.4.4, tab. 6]. The model has 198.7 M parameters and 124.3
GFLOPs, and runs at about 10 FPS with I/O included, on hardware the paper does
not state [arxiv:2204.05585 §4.4.5, tab. 7]. Card:
[SwinNet: Swin Transformer drives edge-aware RGB-D and RGB-T salient object detection](/papers/2204.05585).

### Eye-fixation prediction: DVA

Deep Visual Attention Prediction (DVA) predicts where people look, not which
object is salient. It is included here because its output is also a per-pixel
map [arxiv:1705.02544 §1]. It uses a VGG-16 encoder with the final pooling
layer removed, and three deconvolution decoders on conv3-3, conv4-3 and
conv5-3. Each side output has its own cross-entropy loss, and a learned fusion
layer merges the three maps [arxiv:1705.02544 §3.1, §3.2, §3.3.2]. It is
trained on 10,000 SALICON images only [arxiv:1705.02544 §3.3.3].

Scores use fixation metrics, not the mask measures above: NSS, CC, SIM, EMD
and the AUC variants [arxiv:1705.02544 §4.2]. Author-reported: best on most
metrics for MIT1003, TORONTO, PASCAL-S and DUT-OMRON, but not on MIT300, where
DeeFix and SALICON score higher [arxiv:1705.02544 §4.3, tab. 4, tab. 5]. The
DeeFix row on MIT1003 is fine-tuned on that dataset and is not comparable with
the other rows [arxiv:1705.02544 tab. 5]. Runtime is about 0.1 s per image on
one GPU [arxiv:1705.02544 §3.3.3]. Card:
[Deep Visual Attention Prediction](/papers/1705.02544).

### Video saliency: dynamic FCN

Wang et al. extend the per-frame model to video with two networks
[arxiv:1702.00871 §3.1]. The static network scores one frame. The dynamic
network reads frame t, frame t+1 and the static map of frame t, and outputs
the map for frame t. It computes no optical flow at test time, and the authors
argue that flow, used by earlier video methods, dominates their running time
[arxiv:1702.00871 §1, §3.3, §5.4]. The text does not say what happens to the
last frame, which has no successor [arxiv:1702.00871 §3.3].

The data is the hard part. Labelled video is small, so each annotated still
image from MSRA10K and DUT-OMRON is turned into a two-frame training video,
with ten simulated frames per image [arxiv:1702.00871 §4, §5.1]. The simulated
frame is a warp of the image under a flow field built on superpixels
[arxiv:1702.00871 §4, eq. 5]. Training then mixes these synthetic clips with
about 3K real frame pairs from SegTrackV2 and the FBMS training set
[arxiv:1702.00871 §5.1].

Author-reported MAE, in percent (lower is better) [arxiv:1702.00871 tab. 2]:
7.65 on the FBMS test set and 6.36 on DAVIS for the full model. Training on
image-derived video alone gives 9.27 and 7.53; video alone gives 24.5 and 23.9.
Removing the static prior from the dynamic network gives 9.43 and 8.32
[arxiv:1702.00871 tab. 2]. The paper plots its baselines and gives no numbers
for them, so the card cannot back a numeric comparison [arxiv:1702.00871 fig. 7].
Because the metric is MAE in percent on video test sets, these values are not
comparable with the image MAE values above or with S-measure. Speed is about
2 fps with all steps, on the authors' hardware [arxiv:1702.00871 §5.4].
Card: [Video Salient Object Detection via Fully Convolutional Networks](/papers/1702.00871).

### Concealed-object network: SINet-V2

SINet-V2 predicts a binary mask for concealed objects, a class-agnostic task
[arxiv:2102.10274 §2]. It follows a search-then-identify pattern: a neighbor
connection decoder finds a coarse region, and then three group-reversal
attention (GRA) blocks per stage refine it [arxiv:2102.10274 §4.1]. The
encoder is Res2Net-50, and only the top three feature levels are used
[arxiv:2102.10274 §4.2]. Each stage reverses the coarser prediction, so the
refinement targets the area the previous stage has not yet covered
[arxiv:2102.10274 §4.3, eq. 2].

Training uses Adam with learning rate 1e-4, batch size 36 and 100 epochs at
352 × 352 input, with the rate divided by 10 every 50 epochs. The authors
report about four hours on one TITAN RTX [arxiv:2102.10274 §4.4.2]. Inference
runs once with no post-processing, at about 45 fps on one GPU excluding I/O
[arxiv:2102.10274 §4.4.2].

Author-reported S-measure on three test sets [arxiv:2102.10274 tab. 2]:
CHAMELEON 0.888, CAMO 0.820, COD10K 0.815. The next best S-measure on
COD10K is PraNet's 0.789 [arxiv:2102.10274 tab. 2]. Its dataset, COD10K
(10,000 images), is described in [arxiv:2102.10274 §3]. Card:
[Concealed Object Detection](/papers/2102.10274).

### Camouflage-gated segmentation: ANet

ANet pairs a saliency-style segmentation FCN with a classification stream
[arxiv:2105.09451 §4.2]. The classifier has three fully connected layers,
the first two with ReLU and 50% dropout, and a two-way softmax
[arxiv:2105.09451 §4.2, tab. 1]. Its probability for "camouflaged" is a
single scalar, and it multiplies every pixel of the segmentation map
[arxiv:2105.09451 §4.2]. The authors use four FCNs: DHS, DSS, SRM and WSS
[arxiv:2105.09451 §5.3]. Their names are only pointers here; their methods
have no cards.

Training is two-step. The segmentation stream is first fine-tuned on CAMO,
then frozen while the classifier is trained on CAMO-COCO, which adds
non-camouflaged images [arxiv:2105.09451 §5.1, §5.3]. A joint variant that
sums both losses is reported as mixed [arxiv:2105.09451 §5.5.2, tab. 6].

Author-reported, with an adaptive threshold (mean plus standard deviation)
[arxiv:2105.09451 §5.2, tab. 3]:

| Method | CAMO F | CAMO IoU | CAMO-COCO F | CAMO-COCO IoU |
|---|---|---|---|---|
| SRM (FT CAMO) | 0.683 | 0.507 | 0.815 | 0.651 |
| ANet-SRM (baseline) | 0.654 | 0.475 | 0.826 | 0.732 |

The paper's classifier accuracy on CAMO-COCO is 89.2–91.4% across the four
ANet variants [arxiv:2105.09451 tab. 4]. Adding the classifier costs 3 to
10 ms per image on the paper's hardware [arxiv:2105.09451 tab. 5]. The
authors expect ANet to be comparable with the FCN fine-tuned on CAMO, and
say it can beat that FCN on CAMO-COCO [arxiv:2105.09451 §5.4.2]. Card:
[Anabranch Network for Camouflaged Object Segmentation](/papers/2105.09451).

### Camouflage with difficulty ranks: COD-Rank

Lv et al. use one ResNet-50 network with three outputs: a fixation map for the
discriminative region, a camouflaged-object mask, and a difficulty rank for
each object [arxiv:2103.04011 §3.2, §3.3]. The fixation branch's output gives
a reverse attention, 1 − F, which guides the mask branch [arxiv:2103.04011
§3.2]. The rank labels come from eye-tracker detection delays: six observers
per image on the CAM-FR set of 2,000 training and 280 test images
[arxiv:2103.04011 §3.1]. The paper gives the three rank levels but not the
cut points between them [arxiv:2103.04011 §3.1].

Author-reported, with the joint model trained on COD10K plus CAMO (4,040
images), COD10K S-measure is 0.793 against 0.776 for SINet on the same
training set [arxiv:2103.04011 tab. 4]. On the smaller CAM-FR training set the
joint model scores COD10K S-measure 0.760 against 0.733 for SINet
[arxiv:2103.04011 tab. 1]. Its rank MAE is 0.049 jointly and 0.046 for a
ranking-only model [arxiv:2103.04011 tab. 5], so the joint model's ranking
advantage is not shown. The released repository has no ranking branch (see
the card's Paper vs code section). Card:
[Simultaneously Localize, Segment and Rank the Camouflaged Objects](/papers/2103.04011).

## When to use which

| Situation | Use | Why |
|---|---|---|
| Reporting SOD results that should reflect object shape | S-measure | Best on MM1 and MM3 on all four datasets, and preferred by viewers over Fbw, AP and AUC picks [arxiv:1708.00786 tab. 1, §5.6] |
| Downstream use is a segmentation step such as SalCut | S-measure | Its ranking agrees best with the application ranking (MM1) [arxiv:1708.00786 §5.1, tab. 1] |
| Robustness to slightly edited GT masks matters most | No clear winner | S-measure is not best on MM4 (morphological GT edits), though it beats AP and AUC; the authors consider MM4 unreliable [arxiv:1708.00786 §5.4] |
| Comparing with numbers in older papers | Report AP/AUC/Fbw too | The paper still compares against them, and S-measure ties or loses on some cells (SOD and ECSSD MM2) [arxiv:1708.00786 tab. 1] |
| Need numbers comparable with the official toolkit | Port the official code exactly | Paper text and code differ in block weights and λ; see the card's Paper vs code section [arxiv:1708.00786 §4.1, §5] |
| Object outlines matter and a ResNet or VGG backbone is available | EGNet | Author-reported: best max F, MAE and S-measure of the 15 compared methods on all six test sets, with the ResNet backbone [arxiv:1908.08297 tab. 2, §4.4] |
| Depth map is available but may be poor | D3Net | The gate falls back to the RGB prediction when depth disagrees; author-reported S-measure 0.900 on NJU2K test against 0.879 for CPFP [arxiv:1907.06781 §4.2, tab. 4] |
| Depth or thermal input is available and about 199 M parameters is acceptable | SwinNet | Author-reported best S-measure on NLPR and NJU2K among the compared RGB-D methods, and on VT5000 among the RGB-T baselines [arxiv:2204.05585 §4.3, tab. 1, tab. 2]; cost in [arxiv:2204.05585 §4.4.5, tab. 7] |
| Depth map quality is uncertain | Check RGB alone first | On STERE, RGB alone scores S 0.923 against 0.919 for RGB plus depth in SwinNet's ablation [arxiv:2204.05585 tab. 6, §4.4.4]; D3Net's gate is another way to fall back to RGB [arxiv:1907.06781 §4.2] |
| The target is where people look, not an object mask | DVA | Author-reported best on most metrics for four of five eye-tracking sets, not MIT300; trained on SALICON only [arxiv:1705.02544 §3.3.3, §4.3] |
| Input is video and per-frame speed matters | Dynamic FCN with a static prior | Skips optical flow; author-reported about 2 fps with all steps, and timing is compared with four video methods [arxiv:1702.00871 §5.4, fig. 9] |
| Labelled video is scarce but annotated still images exist | Synthetic video from images, mixed with real pairs | Image-only training raises FBMS MAE from 7.65 to 9.27, so the synthetic data is not a full substitute [arxiv:1702.00871 tab. 2, §5.3] |
| Output must be causal, with no look-ahead frame | Not the dynamic FCN | It reads frame t+1 to produce the map for frame t [arxiv:1702.00871 §3.3] |
| Target is camouflaged and a single pass with no post-processing is wanted | SINet-V2 | Author-reported best S-measure on CHAMELEON, CAMO and COD10K against 12 baselines; one pass at about 45 fps on one GPU [arxiv:2102.10274 tab. 2, §4.4.2] |
| Target is camouflaged and it may be absent from some images | ANet | The classifier decides whether there is any camouflage to segment [arxiv:2105.09451 §4.1, §4.2]; the authors report 89–91% classifier accuracy [arxiv:2105.09451 tab. 4] |
| Need a per-object difficulty rank for camouflaged objects | COD-Rank, with the ranking branch written by you | Ranks come from eye-tracker delays, but the paper gives no cut points and the released code has no ranking branch [arxiv:2103.04011 §3.1, §3.3] |
| Every image contains camouflage and the FCN can be fine-tuned on it | A fine-tuned FCN without the gate | The ANet gate is designed for images that may lack camouflage [arxiv:2105.09451 §4.1]; on CAMO, ANet-SRM's F-beta (0.654) is below its fine-tuned SRM (0.683) [arxiv:2105.09451 tab. 3] |
| Comparing EGNet's S-measure with Structure-measure numbers | Check the implementation first | Both use γ = α = 0.5 [arxiv:1908.08297 eq. 13] [arxiv:1708.00786 §4.3]; the same value does not guarantee the same code |
| Reproducing EGNet's reported numbers from the paper alone | Expect gaps | The optimizer, the fusion weights β_i, input size and edge-label generation are not stated [arxiv:1908.08297 §3.3, §4.1]; the paper gives momentum and weight decay without an optimizer |
| Reproducing SINet-V2 | Pick a validation split | The released code picks the best epoch on the CAMO test set by default, which is one of the reported benchmarks (see the card's Pitfalls) [arxiv:2102.10274 §5.1] |
| Reproducing D3Net | Pick a validation split and check the gate | The released code picks its checkpoint on the NJU2K test split, a reported benchmark; the threshold t = 0.15 was chosen on data the paper does not name [arxiv:1907.06781 §4.3] |
| Reproducing SwinNet from the paper alone | Expect gaps | The paper does not say which Swin outputs form its four feature levels, and does not state the edge-feature or decoder widths [arxiv:2204.05585 §3.2, §3.4, §3.5]; gradient clipping is not stated either [arxiv:2204.05585 §4.2] |
| Reproducing ANet | Reimplement from the paper, do not copy | The paper leaves the classifier's input layer, its fusion point and its widths unresolved [arxiv:2105.09451 §4.2, tab. 1]; the released repository lacks the network definitions (see the card's Paper vs code) |
| Redistributing CAMO images or annotations | Check the dataset's own terms | The CC BY licence covers the paper; the paper states no licence for the images, which were collected by keyword search [arxiv:2105.09451 §3.2] |
| Latency or memory budget is tight | No like-for-like cost data in this wiki | The EGNet paper's timing claim ("no additional time and space consumption") is not backed by numbers in its text [arxiv:1908.08297 §4.3.2, §4.4]; D3Net runs three VGG-16 streams at test time [arxiv:1907.06781 §7.2]; SwinNet reports 198.7 M parameters and about 10 FPS with I/O included, on unstated hardware [arxiv:2204.05585 §4.4.5, tab. 7]; the other papers time on different hardware |
| The model must be built from released code | Use the licence check first | EGNet's, D3Net's, DVA's and COD-Rank's repositories have no LICENSE file, so their code cannot be reused under the wiki's licensing rules; SINet-V2's code is Apache-2.0 (see its card) |

## Implementation sketch

Four cores. The first is the S-measure as its card documents it. Where the
paper and the official code differ, the comment says which one is shown.

```text
input: SM in [0,1] (float), GT binary, both H×W
m = mean(GT)
if m == 0: return 1 - mean(SM)          # code only, not in paper
if m == 1: return mean(SM)              # code only, not in paper

# object term
O_FG = 2·mean(SM[GT]) / (mean(SM[GT])² + 1 + 2λ·std(SM[GT]))
O_BG = same formula on (1 - SM)[not GT]
S_o  = m·O_FG + (1 - m)·O_BG

# region term
(X, Y) = rounded centroid of GT foreground
split SM and GT into 4 blocks at (X, Y)
for each block k: ssim_k = 4·x̄·ȳ·σ_xy / ((x̄² + ȳ²)(σ_x² + σ_y²))   # no C1, C2
w_k = foreground share of block k (paper) | pixel share of block k (code)
S_r = Σ w_k · ssim_k

S = 0.5·S_o + 0.5·S_r                    # code clamps negative S to 0
```

Implementation details that change the number: sample (N − 1) standard
deviations and variances in the code, 1-based rounded centroid indexing,
guarding empty blocks, and not substituting a library SSIM with constants
and sliding windows [arxiv:1708.00786 §4.1]. The card's Pitfalls section
lists each with code line references.

The second is the edge-guided core of EGNet. Shapes and the upsampling
operator are simplified; the card has the full layer table.

```text
input: image; side features C2..C6 from the backbone
# object path, top-down                     [arxiv:1908.08297 §3.2.1, eq. 3]
F6 = enhance(C6)
for i in 5..3: F_i = enhance(C_i + up(F_{i+1}, C_i))
# edge path: C2 plus top-level location     [arxiv:1908.08297 §3.2.2, eq. 2]
F_E = enhance(C2 + up(F6, C2))
edge_map = head(F_E)                        # trained with edge loss, eq. 4
# guide each object scale with edges        [arxiv:1908.08297 §3.3, eq. 7]
for i in 3..6: G_i = enhance'(up(F_i, F_E) + F_E); map_i = head'(G_i)
output = Σ β_i · map_i                      # β_i not stated, eq. 9
```

The third is D3Net's test-time depth gate. The paper's test is "at most t";
the released code uses a strict "below t", which the card records.

```text
input: s_rgb, s_rgbd, s_depth in [0,1], H×W
# mean absolute disagreement between RGB-D and depth-only maps  [arxiv:1907.06781 §4.2, eq. 2]
d = mean(|s_rgbd - s_depth|)
# keep the RGB-D map only when the two agree within t = 0.15  [arxiv:1907.06781 §4.2, eq. 3, §4.3]
return s_rgbd if d <= t else s_rgb
```

The fourth is ANet's presence gate. Only the multiplication is shown; the
fusion point and the classifier's input layer are not stated in the paper.

```text
input: image; seg_map = FCN(image) in [0,1], H×W
# presence stream: three FC layers on FCN features  [arxiv:2105.09451 §4.2, tab. 1]
p_camo = softmax(fc3(drop(relu(fc2(drop(relu(fc1(feat))))))))[camo]
# fusion: scale every pixel by the image-level probability  [arxiv:2105.09451 §4.2]
out = p_camo · seg_map                      # which map and which index: not stated
# scoring: threshold mean + std, or fixed 0.5  [arxiv:2105.09451 §5.2]
```

The unstated values listed above are the ones an implementer has to choose
and then report.

## Contradictions

- **λ inside the S-measure paper:** the Fig. 2 caption gives λ = 0.25, the
  experiments use λ = 0.5 [arxiv:1708.00786 fig. 2, §5].
- **Block weights:** the text weights blocks by the GT foreground area they
  contain [arxiv:1708.00786 §4.1]; the card reports that the official code
  weights them by pixel area instead.
- **Edge loss within the EGNet source:** the paper's edge loss is plain
  cross-entropy over pixels [arxiv:1908.08297 eq. 4]. The EGNet card records
  that the released code weights edge and background pixels differently. The
  paper does not say whether its reported results used the balanced or the
  plain loss [arxiv:1908.08297 §4.3].
- **Size of the CAMO dataset:** SINet-V2 says CAMO has 2,500 images
  [arxiv:2102.10274 §2]; ANet says CAMO has 1,250 images, split 1,000 for
  training and 250 for testing [arxiv:2105.09451 §3.2, §5.1]. The two papers
  use the same name, so numbers on "CAMO" cannot be compared until the sizes
  are resolved. Neither card resolves it.
- **ANet-SRM on CAMO, two F values:** SINet-V2 reports weighted F 0.484 for
  ANet-SRM on CAMO, with MAE 0.126 [arxiv:2102.10274 tab. 2]. ANet reports
  F-beta 0.654 with the adaptive threshold, and MAE 0.126 [arxiv:2105.09451
  tab. 3]. The MAE matches; the F values differ in metric and threshold, and
  the cards do not say which setting SINet-V2 used.
- **The same baselines on COD10K, different numbers:** SINet-V2 reports
  EGNet at S-measure 0.737 and the conference SINet at 0.771 on COD10K
  [arxiv:2102.10274 tab. 2]. COD-Rank reports EGNet at 0.751 and SINet at 0.776
  on COD10K [arxiv:2103.04011 tab. 4]. Both papers retrained these baselines
  on their own training data (SINet-V2 on combined CAMO and COD10K training
  sets [arxiv:2102.10274 §5.1]; COD-Rank on 3,040 COD10K and 1,000 CAMO
  images [arxiv:2103.04011 tab. 4]). The gap is therefore a protocol difference,
  and neither side is wrong about its own table.
- **Different metrics and test sets:** SINet-V2 reports S-measure, E-measure,
  weighted F and MAE on CHAMELEON, CAMO and COD10K [arxiv:2102.10274 §5.1];
  ANet reports F-beta, IoU and MAE on CAMO and CAMO-COCO [arxiv:2105.09451
  §5.2]; COD-Rank reports S-measure, F-beta, E-measure and MAE on CAMO,
  CHAMELEON, COD10K and NC4K [arxiv:2103.04011 §4.1]; DVA uses fixation
  metrics on five eye-tracking sets [arxiv:1705.02544 §4.2]; the video model
  reports MAE in percent on DAVIS and FBMS [arxiv:1702.00871 §5.2]. Apart from
  the ANet-SRM MAE above and the shared COD10K baselines, no two papers share a
  metric and a test set, so the families cannot be ranked against each other.
- **Video MAE, paper against released code:** the paper reports 6.36% on DAVIS
  and 7.65% on FBMS [arxiv:1702.00871 tab. 2]. The card records that the
  authors' repository README reports 0.0587 on DAVIS and 0.0619 on FBMS for its
  Keras code, and calls these slightly better. The released code is a Keras
  re-implementation, while the paper says Caffe [arxiv:1702.00871 §5.1]. Neither
  card resolves which numbers match the paper's model.
- **RGB-D numbers from D3Net and SwinNet:** D3Net trains on 1,485 NJU2K and 700
  NLPR pairs [arxiv:1907.06781 §4.3]. SwinNet's RGB-D training uses the same
  2,185 pairs, and can add DUT for 2,985 [arxiv:2204.05585 §4.1.1]. D3Net
  reports S-measure 0.900 on NJU2K and 0.912 on NLPR [arxiv:1907.06781 tab. 4];
  SwinNet reports 0.935 and 0.941 [arxiv:2204.05585 tab. 1]. The cards do not
  show whether the test splits and evaluation code match, so the gap is not
  resolved here.

## Papers

- [Structure-measure: A New Way to Evaluate Foreground Maps](/papers/1708.00786): structure-aware score for non-binary saliency maps; compares it with AP, AUC and Fbw on five datasets.
- [EGNet: Edge Guidance Network for Salient Object Detection](/papers/1908.08297): object and edge paths over one backbone, with edge features guiding each object scale; author-reported max F-measure 0.893 on DUTS-TE with ResNet.
- [Rethinking RGB-D Salient Object Detection: Models, Data Sets, and Large-Scale Benchmarks](/papers/1907.06781): three-stream RGB-D network with a test-time depth gate; adds the SIP phone-captured dataset and an 18-model benchmark.
- [Deep Visual Attention Prediction](/papers/1705.02544): VGG-16 with three deconvolution decoders for eye-fixation maps; trained on SALICON, author-reported best on most metrics for four of five eye-tracking sets.
- [Concealed Object Detection](/papers/2102.10274): SINet-V2, a search-then-identify network for concealed objects with reverse-guided refinement; author-reported S-measure 0.815 on COD10K.
- [Anabranch Network for Camouflaged Object Segmentation](/papers/2105.09451): a camouflage classifier scales a saliency FCN's map; introduces the CAMO dataset; author-reported F-beta 0.826 on CAMO-COCO with ANet-SRM.
- [Simultaneously Localize, Segment and Rank the Camouflaged Objects](/papers/2103.04011): one network for fixation, camouflage masks and difficulty rank; the ranking branch is not in the released code.
- [Video Salient Object Detection via Fully Convolutional Networks](/papers/1702.00871): static and dynamic FCNs for video, with no optical flow; synthetic video from annotated images; author-reported MAE 7.65% on FBMS and 6.36% on DAVIS.
- [SwinNet: Swin Transformer drives edge-aware RGB-D and RGB-T salient object detection](/papers/2204.05585): two Swin-B streams with per-level alignment and an edge-guided decoder; author-reported S-measure 0.941 on NLPR and 0.912 on VT5000.