Weakly-Supervised Camouflaged Object Detection with Scribble Annotations

Introduction

• Scribble-based COD addresses costly pixel-wise labeling, which takes about 60 minutes per image compared with about 10 seconds for a scribble.The S-COD dataset comprises 3,040 COD10K training images and 1,000 CAMO training images.Source: Introduction · p. 1
• Existing scribble-based salient object detection methods struggle with COD boundaries and semantic relations, confusing lookalike objects with foreground.Ambiguous boundaries in consistent backgrounds challenge low-level cues, while related objects such as flower stems and petals require semantic reasoning.Source: Introduction · p. 2
• Feature-guided loss combines visual affinity with learned high-level semantic features to guide segmentation without relying on an external pretrained detector.The learned semantic features are trained end-to-end.Source: Introduction · p. 2
• The framework combines cross-view and inside-view consistency losses with LCC contrast enhancement and LSR semantic reasoning to expand scribbles and stabilize predictions.Cross-view loss favors more reliable predictions across input transformations, while inside-view loss reduces prediction uncertainty.Source: Introduction · p. 2
• On three COD benchmarks, the framework outperforms relevant state-of-the-art methods, achieving average improvements of 11.0% on MAE, 3.2% on S-measure, 2.5% on E-measure, and 4.4% on weighted F-measure.Source: Introduction · p. 2

Related Work

• Prior COD methods use graph learning, joint SOD-COD learning, focus modules, probabilistic representations, and frequency-aware or multi-scale networks.These approaches target rough positioning and boundary localization, global/local information, false predictions, uncertainty, and mixed-scale semantics.Source: Related Work · pp. 2–3
• These methods rely heavily on fully supervised per-pixel ground truth, which is time-consuming and labor-intensive..Source: Related Work · p. 3
• The paper proposes scribble annotations as the first weakly supervised dataset approach for COD, to the authors’ knowledge.Source: Related Work · p. 3

Methodology

• Each training example pairs an input image with a scribble map, where 1 marks foreground, 2 background, and 0 unknown pixels.Source: Methodology · p. 3

Overall Structure

• CRNet uses ResNet-50 features, LCC on low-level stages, LSR on high-level stages, and AGE for global semantics before producing main and auxiliary maps.The intermediate feature map Fss is also extracted for loss computation.Source: Overall Structure · p. 3

Local-Context Contrasted (LCC) Module

• LCC contrasts local and context features from different receptive fields, strengthening low-level differences so scribbles can expand toward potential object regions and boundaries.Its two extractors use different receptive fields to represent local and neighboring context features.Source: Local-Context Contrasted (LCC) Module · p. 3

Logical Semantic Relation (LSR) Module

• LSR integrates semantic features from four branches with different receptive fields to distinguish foreground from background when scribbles label only part of the background.The branches use different kernel sizes and dilation rates.Source: Logical Semantic Relation (LSR) Module · p. 3

Feature-guided Loss

• Feature-guided supervision pairs visual context affinity with semantic-significance guidance, using semantic features to refine predictions specifically in boundary regions.Context affinity uses visual features such as color and position; semantic significance uses features learned by the model.Source: Feature-guided Loss · pp. 3–4

Consistency Loss

• Cross-view consistency aligns normal and transformed predictions toward the more reliable normal view, while inside-view consistency favors confident predictions and filters uncertain, noisy pixels.The cross-view transform may combine resizing, flipping, translation, and cropping.Source: Consistency Loss · p. 5

Objective Function

• The total objective combines consistency, feature-guided, and scribble-label PCE losses; auxiliary maps receive PCE, inside-view, and context-affinity losses.The loss terms are applied to the main output, with the specified subset also applied to auxiliary outputs.Source: Objective Function · p. 6

Datasets and Implementation Details

• 4,040 images—3,040 from COD10K and 1,000 from CAMO—were relabeled with scribbles for S-COD training; the remaining images were used for testing.Evaluation used CAMO, CHAMELEON, and COD10K, with MAE, S-measure, E-measure, and weighted F-measure.Source: Datasets and Implementation Details · p. 6
• Training used 320×320 images, horizontal flips, SGD, batch size 16, and 150 epochs, taking about five hours; inference resizes images without post-processing.The experiments were implemented in PyTorch on a GeForce RTX2080Ti GPU.Source: Datasets and Implementation Details · p. 6

Analysis

• 11.0% lower MAE, 3.2% higher S-measure, 2.5% higher E-measure, and 4.4% higher weighted F-measure than SCWSSOD on average; the method also outperforms seven fully supervised methods.It performs best among weakly or unsupervised methods under all four metrics on all three benchmarks.Source: Analysis · p. 6
• The method performs best on CAMO relative to the other two benchmarks, and Figure 6 compares methods in challenging scenarios including tiny objects and complex backgrounds.On CAMO, LCC identifies potential regions but can be confused by complex backgrounds, while LSR can produce inaccurate boundaries; combining them improves performance over either module alone.Source: Analysis · p. 7, Figure 6 · p. 7
• Flipping, translating, and cropping significantly improved cross-view consistency, while consistency loss improved all metrics except MAE and both proposed losses substantially improved test results.Table 3 groups ablations by cross-view transformations, consistency loss, feature-guided loss, and their combined use.Source: Analysis · p. 7, Table 3 · p. 7

Conclusion

• Scribble annotations take about 10 seconds per image, 360 times faster than pixel-wise labeling, and avoid assigning equal importance to every object pixel.The proposed framework combines consistency losses, feature-guided losses, and a network that expands scribbles using low-level contrast before using semantic relations to identify foreground and background.Source: Conclusion · p. 7

• The author list names Ruozhen He, Qihua Dong, Jiaying Lin, and Rynson W.H. Lau.Source: Supplementary Material:… · p. 10
• The metadata identifies the paper as arXiv:2207.14083v2, dated 28 Nov 2022.Source: Supplementary Material:… · p. 10

Related Work

• Scribble-based salient-object methods use edge detection, local coherence, and consistency constraints, but camouflage’s subtle appearance can mislead local-coherence cues and makes those methods generally unsuitable for COD.The cited approaches include Zhang et al. 2020b and Yu et al. 2021.Source: Related Work · p. 10

Weakly Supervised Salient Object Detection

• Image-level salient-object methods often derive pseudo-labels using CAM, which distinguishes only the most distinctive class regions, while hand-crafted alternatives may not exploit rich data information.The passage cites Zhou et al. 2016, Li, Xie, and Lin 2018, and Zeng et al. 2019.Source: Weakly Supervised Salient Object… · p. 10
• Point-level methods train in either two-stage or one-stage fashion, and one-stage approaches gained popularity after gated CRF loss enabled comparable performance with greater simplicity and efficiency.The passage cites Obukhov et al. 2019 for gated CRF loss and lists point-level methods including Vernaza and Chandraker 2017, Piao et al. 2022, Gao et al. 2022, Yu et al. 2021, and Zhang et al. 2020.Source: Weakly Supervised Salient Object… · p. 10

S-COD Dataset

• S-COD contains 4,040 images of diverse objects in challenging scenarios, including similar appearances, partial occlusions, and multiple objects.Source: S-COD Dataset · p. 10
• Annotators mark objects’ primary structures in scribbles based on first impressions, without knowing the ground truth.Source: S-COD Dataset · p. 10
• The dataset addresses pixel-wise annotation’s time-consuming and labor-intensive nature.Source: S-COD Dataset · p. 10

Local-Context Contrasted (LCC) Module

• LCC contrasts local and contextual features, using context dilation rates of 4 and 8 to capture different receptive-field sizes.Each low-level contrasted extractor combines a local receptor, a context receptor, and low-level feature extractors.Source: Local-Context Contrasted (LCC) Module · p. 10
• LCC’s low-level feature extractor uses horizontal and vertical global pooling to create attention-weighted features from texture, intensity, and color.The pooled features are processed and multiplied with the input feature.Source: Local-Context Contrasted (LCC) Module · p. 11

Logical Semantic Relation (LSR) Module

• LSR analyzes logical semantic information with four convolutional branches, concatenating their outputs before a residual addition to the input feature.The branches use different convolutional sequences, including dilated 3×3 layers.Source: Logical Semantic Relation (LSR) Module · pp. 11–12

Limitations and Future Work

• Tiny, dense grasses occluding a cat are missed even when the cat’s primary structure is predicted.Source: Limitations and Future Work · p. 12
• For an animal with slender tentacles, the method predicts the main body and some tentacles near it, but not the entire tentacles.Source: Limitations and Future Work · p. 12
• Future work will focus on local details, especially intricate boundaries, to improve segmentation precision under weak supervision.Source: Limitations and Future Work · p. 12