Abstract
Camouflage Object Detection (COD) is a challenging task due to the high similarity between targets and their backgrounds. While Vision Transformers (ViTs) have set new benchmarks, their computational cost is significant. This paper introduces CamoFocus, an efficient framework that utilizes "Focal Modulation" to explicitly separate foreground and background features. By refining these features across multiple scales, CamoFocus achieves state-of-the-art performance with a leaner parameter footprint.
Introduction
Detecting camouflaged objects is often described as a "needle in a haystack" problem. Unlike generic object detection, camouflaged targets are naturally or artificially designed to blend seamlessly with their environment's texture and color. Existing state-of-the-art (SOTA) methods frequently rely on complex attention mechanisms which, while powerful, often struggle with fine-grained structural details and carry high computational overhead. CamoFocus addresses these challenges by rethinking feature interaction through a split-stream modulation approach.
Methodology
The CamoFocus architecture is designed around three primary components that prioritize efficiency and structural clarity.
Feature Split and Modulation (FSM)
The core innovation of CamoFocus is the FSM module. Rather than processing feature maps uniformly, FSM uses a predicted supervisory mask to bifurcate features into Foreground and Background streams. Each stream is processed by dedicated modulators (FM and BM) using depth-wise convolutions with varying kernel sizes. This explicit separation prevents "feature washing," where dominant background noise might otherwise obscure subtle object details.
Context Refinement Module (CRM)
To handle objects of varying scales, the CRM fuses multi-scale features. It employs dilated convolutions and skip connections to capture a wider receptive field without significantly increasing the parameter count. This ensures that both global context and local details are preserved.
Hybrid Loss Function
Training is guided by a combination of Weighted Binary Cross-Entropy (BCE), Weighted Intersection Over Union (IoU), and Dice Loss. This hybrid approach forces the model to focus on challenging boundary pixels and maintain the overall structural integrity of the detection.
Experiments and Results
CamoFocus was evaluated on four major COD benchmarks: CAMO, COD10K, NC4K, and CHAMELEON.
Performance Metrics
On the NC4K dataset, CamoFocus achieved a Structure-Measure ($S_m$) of 0.889. Notably, it outperformed larger models like CamoFormer-P4 while utilizing a more compact backbone. This demonstrates that architectural innovation in feature modulation can be more effective than simple scaling.
Qualitative Analysis
Visual results indicate that CamoFocus excels at recovering sharp boundaries, particularly for thin or highly obscured objects such as insect legs or sticks, which competing models often blur into the background.
Conclusion
CamoFocus presents a compelling alternative to attention-heavy COD models. By explicitly splitting foreground and background features and applying focal modulation, it achieves superior accuracy and efficiency. Future work may address the model's dependency on the initial mask generation and explore its application in video-based camouflage detection.
Summary updated