Source-linked AI summary
Boundary-Guided Camouflaged Object Detection
Yujia Sun, Shuo Wang, Chenglizhao Chen, Tian-Zhu Xiang
TL;DR
Camouflaged object detection must identify objects that closely blend into their surroundings while preserving complete and fine structure. BGNet uses object-related edge semantics to guide representation learning, and experiments show stronger performance than existing state-of-the-art methods across three benchmarks.
Problem
Camouflaged object detection is challenging because objects blend with their surroundings, making complete and fine structural identification difficult.
Method
BGNet extracts object-related edge semantics and integrates them with camouflaged-object features to guide representation learning.
Results
BGNet outperforms existing state-of-the-art methods on three benchmark datasets.
Takeaways & Limitations
Using edge cues enables BGNet to produce camouflaged-object predictions with complete and fine object structure and boundaries.
Abstract
from arXiv · showhide
Camouflaged object detection (COD), segmenting objects that are elegantly blended into their surroundings, is a valuable yet challenging task. Existing deep-learning methods often fall into the difficulty of accurately identifying the camouflaged object with complete and fine object structure. To this end, in this paper, we propose a novel boundary-guided network (BGNet) for camouflaged object detection. Our method explores valuable and extra object-related edge semantics to guide representation learning of COD, which forces the model to generate features that highlight object structure, thereby promoting camouflaged object detection of accurate boundary localization. Extensive experiments on three challenging benchmark datasets demonstrate that our BGNet significantly outperforms the existing 18 state-of-the-art methods under four widely-used evaluation metrics. Our code is publicly available at: https://github.com/thograce/BGNet.
1 Introduction
Camouflaged object detection is difficult because objects closely resemble chaotic backgrounds and may have disrupted edges or disguised outlines. BGNet addresses this challenge by using object-related edge semantics to guide feature learning and improve structural and boundary predictions.
- The paper presents edge semantics as an auxiliary cue for distinguishing object structure and producing predictions with intact boundaries.Figure 1 compares BGNet with JCSOD, S-MGL, and R-MGL in challenging scenarios.
- Camouflaged object detection is challenging because candidate objects have high intrinsic similarity to chaotic backgrounds.
- Prior COD methods use targeted architectures, auxiliary tasks, or bio-inspired designs to obtain discriminative features and additional cues.
- Existing methods can produce coarse or incomplete boundaries when edge disruption or disguised body outlines obscure object structure and details.
- BGNet uses an edge-aware module and edge-guidance feature module to extract object-related edge semantics and integrate them with camouflaged-object features.The modules guide representation learning toward object structure and details.
2 Related Work
Related work for camouflaged object detection includes benchmark datasets and several families of deep-learning methods. The field has expanded alongside larger annotated datasets and increasingly diverse model designs.
- Datasets: CAMO contains 1,250 camouflaged images across eight categories, while COD10K contains 10,000 images across 78 categories with hierarchical annotations.
- Datasets: NC4K contains 4,121 images with additional localization and ranking annotations for camouflaged-object evaluation.
- Camouflaged object detection: COD methods are broadly grouped into advanced network architectures, auxiliary-task frameworks, and bio-inspired approaches.
3 Proposed Method
BGNet combines boundary-related edge semantics, multi-level feature guidance, and cross-scale contextual aggregation to improve camouflaged-object representations and predictions. Its training uses separate mask and edge supervisions.
- Overall Architecture: The architecture contains three key components: EAM, EFM, and CAM.EAM excavates edge features, EFM integrates edge cues, and CAM enhances representation with contextual semantics.
- Overall Architecture: BGNet extracts multi-level backbone features, then uses EAM to learn object-related edge semantics from low-level details and high-level location information.EAM combines f2 and f5 under explicit boundary supervision.
- Edge-guidance Feature Module: EFM integrates EAM edge cues with multi-level features to guide representation learning and enhance object-structure semantics.It uses element-wise fusion, skip connections, convolution, and local channel attention to emphasize critical channels and suppress redundant responses.
- Context Aggregation Module: CAM aggregates fused features through cross-scale interaction and atrous convolutions to mine multi-scale contextual semantics for camouflaged-object prediction.The module uses adjacent-branch interaction with dilation rates {1, 2, 3, 4}, followed by concatenation, convolution, and residual processing.
- Training Objective: BGNet applies weighted binary cross-entropy and weighted IoU losses to mask predictions, while Dice loss supervises edge prediction.Mask supervision covers three CAM predictions, and the edge-loss trade-off parameter is set to λ = 3.
4 Experiments
BGNet is evaluated on three public COD benchmarks against 18 state-of-the-art methods using four metrics. It achieves the strongest reported quantitative and qualitative performance, while ablations support the contributions of its modules and edge cues.
- Comparison with State-of-the-arts: BGNet outperforms 18 competing methods on three datasets under four evaluation metrics.The evaluation uses CAMO, COD10K, and NC4K, with Sα, Eφ, F wβ, and M.
- Qualitative Evaluation: Qualitative comparisons show finer, more complete object structures and boundary details across representative COD10K samples.The comparison covers aquatic, terrestrial, flying, and amphibious super-classes.
- Ablation Study: Adding CAM improves the baseline, with a 1.50% average gain on F wβ.
- Ablation Study: The complete BGNet gains 1.10% in Sα, 1.14% in Eφ, and 2.65% in F wβ on average across datasets.The full model combines EAM, EFM, and CAM; EAM uses f2 + f5 for the best EAM input combination.
5 Conclusion
The paper proposes BGNet to use object-related edge semantics for COD representation learning. Across three benchmarks, the method produces accurate predictions with complete, fine structures and boundaries and outperforms existing state-of-the-art methods.
- BGNet combines edge-aware, edge-guidance, and context aggregation modules to guide COD representation learning with object-related edge semantics.
- Across three benchmarks, BGNet outperforms existing state-of-the-art methods while producing complete and fine object structures and boundaries.
6 Appendix
The appendix examines BGNet’s model complexity, component contributions, local channel attention, and loss-weight selection. Results indicate balanced complexity, complementary modules, effective boundary cues, and best performance at λ = 3.
- 6.1 Model complexity: BGNet achieves a good balance of number of parameters, FLOPs, and FPS against multi-task COD methods.The comparison includes S-MGL, R-MGL, JCSOD, and LSR.
- 6.2 Ablation study: Boundary cues remain effective when fused simply, while local channel attention further improves the edge-guidance feature module.The comparison evaluates B+EAM+EFM without and with LCA.
- 6.3 Parameter analysis: λ = 3 achieves the best performance among the tested loss-weight settings on CAMO.The tested values are λ = {1, 2, 3, 4, 5}, evaluated with Eφ, F w β, Sα, and M.
- 6.4 Relationship between modules: EAM, EFM, and CAM are serially interdependent components whose complementary contributions average 1.22%, 0.65%, and 1.94% in F w β, respectively.CAM and EAM contribute more than EFM in the reported analysis.