C3Net: Context-Contrast Network for Camouflaged Object Detection

TL;DR

C3Net addresses camouflaged object detection by introducing a dual-pathway decoder that separately optimizes edge refinement and contextual localization, then fuses them via an attentive mechanism. The Edge Refinement Pathway (ERP) uses gradient- and Laplacian-initialized Edge Enhancement Modules to recover precise boundaries, while the Contextual Localization Pathway (CLP) employs Semantic Enhancement Units and the Image-based Context Guidance mechanism to suppress intrinsic saliency without external models. An Attentive Fusion Module and a triple-loss objective coordinate edge, context, and final predictions, achieving state-of-the-art performance across COD10K, CAMO, and NC4K benchmarks and demonstrating robust handling of six COD challenges. The approach shows that carefully designed specialized pathways and intrinsic saliency suppression can outperform foundation-model baselines, with practical impact for medical imaging, wildlife monitoring, and industrial inspection.

Abstract

Camouflaged object detection identifies objects that blend seamlessly with their surroundings through similar colors, textures, and patterns. This task challenges both traditional segmentation methods and modern foundation models, which fail dramatically on camouflaged objects. We identify six fundamental challenges in COD: Intrinsic Similarity, Edge Disruption, Extreme Scale Variation, Environmental Complexities, Contextual Dependencies, and Salient-Camouflaged Object Disambiguation. These challenges frequently co-occur and compound the difficulty of detection, requiring comprehensive architectural solutions. We propose C3Net, which addresses all challenges through a specialized dual-pathway decoder architecture. The Edge Refinement Pathway employs gradient-initialized Edge Enhancement Modules to recover precise boundaries from early features. The Contextual Localization Pathway utilizes our novel Image-based Context Guidance mechanism to achieve intrinsic saliency suppression without external models. An Attentive Fusion Module synergistically combines the two pathways via spatial gating. C3Net achieves state-of-the-art performance with S-measures of 0.898 on COD10K, 0.904 on CAMO, and 0.913 on NC4K, while maintaining efficient processing. C3Net demonstrates that complex, multifaceted detection challenges require architectural innovation, with specialized components working synergistically to achieve comprehensive coverage beyond isolated improvements. Code, model weights, and results are available at https://github.com/Baber-Jan/C3Net.

Figures (3)

• Figure 1: Visual comparison demonstrating C3Net's comprehensive handling of COD challenges. Each row illustrates a fundamental challenge: (i) Intrinsic Similarity (IS) - a camouflaged insect blends with leaves; (ii) Edge Disruption (ED) - an insect exhibits fragmented boundaries against ground; (iii) Extreme Scale Variation (ESV) - a bear barely visible in dense vegetation occupies minimal pixels; (iv) Environmental Complexities (EC) - shadows and terrain variations obscure half of the ground creature; (v) Contextual Dependencies (CD) - the insect requires global context for accurate segmentation; (vi) Salient-Camouflaged Object Disambiguation (SCOD) - a camouflaged insect must be distinguished from the prominent bark. C3Net consistently outperforms FEDER He_2023_CVPR (CNN-based SOTA) and FSPNet huang2023feature (ViT-based SOTA) across all challenges.
• Figure 2: Overview of the C3Net architecture. Input images are preprocessed and encoded to extract multi-scale features. Our dual-pathway decoder comprises two specialized branches. The Edge Refinement Pathway (top) processes early features through Edge Enhancement Modules (EEMs) to produce detailed edge maps. The Contextual Localization Pathway (bottom) processes deep features through Semantic Enhancement Units (SEUs) and our Image-based Context Guidance (ICG) mechanism to generate object maps with suppressed saliency. The ICG contains three components: appearance analysis from the input image, the Guided Contrast Module (GCM) for foreground-background differentiation, and iterative attention gates for saliency suppression. The Attentive Fusion Module (AFM) combines the outputs of both pathways through spatial gating to produce the final segmentation. Deep supervision is applied at three points: edge map, object map, and final prediction.
• Figure 3: Visual comparison of C3Net with state-of-the-art methods on challenging COD cases. Each row exemplifies a specific challenge: (i) Intrinsic Similarity (IS), (ii) Edge Disruption (ED), (iii) Contextual Dependencies (CD), (iv) Multiple Instances, (v) Environmental Complexities (EC), (vi) Small Objects (ESV aspect), (vii) Large Objects (ESV aspect), and (viii) Salient-Camouflaged Object Disambiguation (SCOD). For each row, columns are: (a) Input Image, (b) OCENet 9706783, (c) BGNet sun2022bgnet, (d) ZoomNet pang2022zoom, (e) SINetV2 fan2021concealed, (f) FSPNet huang2023feature, (g) FEDER He_2023_CVPR, (h) C3Net (Ours), and (i) Ground Truth. Visual results for the compared methods are obtained from officially released predictions.