Advances in Deep Concealed Scene Understanding

Motivation

• CSU recognizes objects that exhibit camouflage and is challenging compared with conventional object detection, with applications in search and rescue, healthcare, agriculture, and content creation.Earlier approaches used hand-engineered patterns, heuristic priors, or combinations of techniques.Source: §1 · p. 1

Recent progress

• Deep learning and the availability of benchmarks such as COD10K and NC4K have accompanied important progress in CSU.COD10K was released in 2020 as the first large-scale public dataset geared toward concealed-object perception tasks.Source: §1 · p. 1

Contributions

• The paper presents a comprehensive survey of deep learning for CSU, covering its taxonomy, task-specific challenges, and developments, alongside extensive benchmarks and discussion of open problems.Its COS benchmark is described as the most up-to-date and comprehensive available, and CDS2K as the largest concealed defect segmentation dataset, with challenging cases from varied industrial settings.Source: §1 · pp. 1–2

2.1.1 Image-level CSU

• Image-level CSU includes COS masks, COL fixation maps, CIR instance camouflage rankings, CIS instance labels, and COC center estimates and density maps.COS, COL, CIR, and CIS are dense prediction tasks, whereas COC is sparse; COS has been studied extensively while research on the other tasks is increasing.Source: §2.1.1 · pp. 2–3

2.1.2 Video-level CSU

• Video-level CSU includes VCOD, which predicts object locations per frame, and VCOS, which predicts per-frame segmentation masks using spatial-temporal cues.Video CSU develops more slowly than image-level CSU; collecting and annotating video data is labor-intensive and time-consuming, though the first large-scale VCOS benchmark brought fundamental progress.Source: §2.1.2 · p. 3

2.1.3 Task Relationship

• CIR requires ranking instances by fixation probabilities and may involve segmentation, counting, instance distinction, and localization; VCOS is downstream of VCOD because segmentation requires pixel-level classification probabilities.Source: §2.1.3 · p. 3

2.2 Related Topics

• Salient object detection extracts pixel-accurate silhouettes of attractive objects, whereas COS focuses on concealed attributes; the tasks are distinct but can mutually benefit through approaches such as CamDiff.Video SOD research also shifted from treating video object segmentation and saliency detection as equivalent toward accounting for attention shifts in multi-object scenes.Source: §2.2 · p. 3

Review organization

• The review systematically organizes deep CSU approaches by task definitions and data types and provides a GitHub collection intended to keep field information current.Source: §3 · p. 3

Image-based methods

• The image-based-methods table summarizes architecture design, auxiliary cues, supervision level, and task level, including COS, CIS, COC, and multi-task learning.Source: Table 1 · p. 4

Further detail

• The section directs readers to §3.1 for more detailed descriptions of image-level CSU model characteristics.Source: §3 · p. 4

3.1.1 Concealed Object Segmentation

• COS methods use multi-stream, bottom-up/top-down, or branched architectures, with branches pairing segmentation and auxiliary tasks.Multi-stream designs learn multi-source representations; bottom-up/top-down designs propagate deeper features to shallower ones, sometimes with deep supervision.Source: §3.1.1 · pp. 4–5, Figure 3 · p. 5
• COS uses single-task segmentation or multi-task learning that complements segmentation with confidence, localization, category, depth, boundary, or texture tasks.Attention modules are common in single-task methods; auxiliary tasks are used to support robust segmentation learning.Source: §3.1.1 · p. 5

3.1.2 Concealed Instance Ranking

• Concealed instance ranking uses localization, segmentation, and ranking together to estimate camouflage difficulty and object detectability.CAM-LDR includes instance masks, fixation labels, and ranking labels; LSR and LSR+ jointly learn the three tasks.Source: §3.1.2 · p. 5

3.1.3 Concealed Instance Segmentation

• Concealed instance segmentation extends COS to the instance level, with approaches including camouflage fusion learning and transformer-based detection.CAMO++ supports the task; OSFormer uses location-sensing and coarse-to-fine fusion, while another approach combines pixel-level decoupling with instance-level suppression.Source: §3.1.3 · p. 5

3.1.4 Concealed Object Counting

• Indiscernible object counting targets objects difficult to distinguish from their surroundings, while proposed solutions remain few for counting, ranking, and localization.IOCfish5K provides high-resolution underwater fish images with dense annotations, and IOCFormer combines density-based and regression-based methods.Source: §3.1.4 · p. 5

3.2.1 Video Concealed Object Detection

• Video concealed object detection is often framed as segmentation under scarce pixel-wise annotations, using offline optical flow and varied real, synthetic, or self-supervision.Methods are commonly trained on segmentation datasets and evaluated for generalization on MoCA.Source: §3.2.1 · p. 6

3.2.2 Video Concealed Object Segmentation

• Video concealed object segmentation has progressed from trajectory-based motion clustering to MoCA-Mask and the two-stage SLTNet baseline.MoCA-Mask adds pixel-level masks to the earlier localization-level MoCA benchmark.Source: §3.2.2 · p. 6
• Current video concealed object segmentation approaches remain nascent, with high-level semantic understanding, including semantic and instance segmentation, identified as gaps.The review points to developments in self-supervised correspondence learning and unified motion-based frameworks as directions for further exploration.Source: §3.2.2 · p. 6

4.1 Image-level Datasets

• CAMO-COCO contains 2,500 images, while NC4K is the largest COS testing set with 4,121 camouflaged images.CAMO-COCO is split between camouflaged CAMO and non-camouflaged MS-COCO images; NC4K includes object-level and instance-level masks.Source: §4.1 · p. 6
• CAMO++ contains 5,500 samples, and COD10K contains 10,000 images spanning ten super-classes and 78 sub-classes.COD10K includes camouflaged, non-camouflaged, and background images, with multiple annotation types for its camouflaged subset.Source: §4.1 · p. 6
• CAM-LDR supplies localization, ranking, object-mask, and instance-mask labels, whereas S-COD provides scribble labels for weakly supervised COS.CAM-LDR uses six ranking difficulty levels; S-COD has 4,040 training samples selected from COD10K and CAMO.Source: §4.1 · pp. 6–7
• IOCfish5K supports concealed object counting with 5,637 high-resolution fish images and 659,024 annotated center points.The review also identifies CAMO, COD10K, and NC4K as common COS benchmarks, and IOCfish5K as a counting dataset.Source: §4.1 · p. 7

4.2 Video-level Datasets

• CAD has nine short clips and sparse masks, while MoCA contains 141 video clips with bounding-box labels for 7,617 instances.MoCA is described as a video database for detecting moving camouflaged animals.Source: §4.2 · p. 7
• MoCA-Mask adds masks every five frames to MoCA, but the review cautions that video annotation quantity and quality remain insufficient for reliable models in complex concealed scenarios.Its training and evaluation sets contain 3,946 and 745 segmentation masks, respectively; pseudo-labels for unlabeled frames use bidirectional optical flow.Source: §4.2 · p. 7

5.1 Evaluation Metrics

• The benchmark compares prediction masks with same-resolution ground truth using MAE, F-measure, structure, and enhanced-alignment measures; F-measure uses β² = 0.3 to emphasize precision.F-measure and enhanced-alignment scores are evaluated over thresholds.Source: §5.1 · pp. 7–8

5.2 Experimental Protocols

• Competing models are trained on 4,040 COD10K and CAMO training samples, then tested on COD10K, CAMO, and NC4K test sets.The test portions contain 2,026 COD10K samples, 250 CAMO samples, and 4,121 NC4K samples.Source: §5.2 · p. 8

5.3 Quantitative Analysis on CAMO

• Among convolutional-backbone models on CAMO, CamoFormer-C with ConvNeXt achieved the best performance, with Sα 0.859 versus DTINet 0.856 and HitNet 0.849.CamoFormer-P achieves comparable results with 71.40M parameters versus 97.27M for CamoFormer-R.Source: §5.3 · p. 9

5.4 Quantitative Analysis on NC4K

• On NC4K, CamoFormer-C remains best; PopNet improves Emnφ by 0.6% over SINetV2 despite a heavier model and larger input.CamoFormer-P performs better than CamoFormer-S on NC4K.Source: Table 5 · p. 9, §5.4 · p. 10

5.5 Quantitative Analysis on COD10K

• On COD10K, HitNet achieves the highest benchmark performance, and the dataset’s higher image resolution suggests high-resolution or multi-scale models may benefit.The authors recommend HitNet where high-resolution images are available.Source: §5.5 · pp. 10–11, Table 6 · p. 10

5.6 Qualitative Comparison

• Models often miss multiple or small objects and fine details; only two of ten locate the flying bird, while only two detect the distant small cat.Big objects are generally detected at their main part, but smaller details can be missed; small objects typically occupy less than 10% of image pixels and may be lost through down-sampling.Source: §5.6 · p. 11
• Out-of-view objects and thin parts remain difficult, most models fail on occluded owls, and indefinable boundaries are hard to resolve.CamoFormer-C and FDNet preserve out-of-view object integrity; only high-resolution HitNet predicts the stick insect’s right-bottom foot, while current models improve on SINet and weakly supervised CRNet locates targets satisfactorily.Source: §5.6 · p. 11

Annotation-Efficient Learning

• Annotation-efficient learning could address the data and resource demands of supervised CSU while supporting generalizability in practical settings.The paper points to weakly supervised learning as one example.Source: §6 · p. 11

Domain Adaptation

• Domain adaptation, transfer learning, few-shot learning, and meta-learning are proposed to address deployment gaps between natural-scene training data and autonomous-driving scenarios.Source: §6 · p. 11

High-Fidelity Synthetic Dataset

• Synthetic data may increase training-data diversity and scale, but models must distinguish camouflaged objects from salient objects in generated samples.The paper cites a diffusion framework that generates salient objects on camouflaged samples while preserving their original labels.Source: §6 · p. 11

Neural Architecture Search

• Neural architecture search could find architectures suited to complex backgrounds, variable object appearances, and limited labeled data, and could be combined with domain adaptation or data-efficient learning.Source: §6 · p. 11

Large Model and Prompt Engineering

• Prompt engineering with robust encoders and task-specific adaptations is proposed for large models, while large language models could assist CSU expert models with auxiliary decisions or representations.The paper notes that SAM has limitations in unprompted concealed scenarios and that training large models is resource-consuming for academia.Source: §6 · pp. 11, 13

7 DEFECT SEGMENTATION DATASET

• CDS2K addresses industrial defects that blend into surrounding materials, which existing vision-based defect methods assume are easy to detect.Source: §7 · p. 13

7.1 Dataset Organisation

• The dataset selects concealed positive defects and contrasting normal negatives from established industrial databases, with selection requiring defects to resemble their background.Sources include MVTecAD, NEU, CrackForest, KolektorSDD, and Magnetic Tile; MVTecAD contributes 748 positive and 746 negative samples.Source: §7.1 · pp. 13–14

7.3 Evaluation on CDS2K

• Current COS models are not effective on positive CDS2K samples, indicating a need to further explore the domain gap between natural scenes and downstream applications.The evaluation tests SINetV2, DGNet, CamoFormer-P, and HitNet.Source: Table 8 · p. 14, §7.3 · p. 14

7.2 Dataset Description

• CDS2K contains 2,492 samples—1,330 positive and 1,162 negative—with human-annotated category, bounding-box, and pixel-wise segmentation labels.The figure illustrates the category and defect-location annotations, including bounding boxes and segmentation masks.Source: §7.2 · p. 14, Figure 6 · p. 14

• The paper provides a comprehensive deep-CSU survey, the largest and most up-to-date COS benchmark, and the largest concealed defect dataset with a practical-scenario benchmark.It also discusses open problems and potential research directions.Source: §8 · pp. 14–15
• Transformer- and ConvNext-based models outperform other competitors by a clear margin, while cost-performance trade-offs remain under-studied.DGNet is cited as an attempt to study the trade-off.Source: §8 · p. 15
• Existing CSU datasets remain insufficiently large and diverse, and the paper reports no well-defined metric that fairly evaluates differing degrees of camouflage.The authors call for more concealed samples from domains such as autonomous driving and clinical diagnosis.Source: §8 · p. 15
• Current CSU methods emphasize appearance over semantic relationships between objects, motivating future work on high-level semantics, vision-language knowledge, and object interactions.Source: §8 · p. 15