Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Shifting Spotlight for Co-supervision: A Simple yet Efficient Single-branch Network to See Through Camouflage

Yang Hu, Jinxia Zhang, Kaihua Zhang, Yin Yuan, Jiale Huang, Zechao Zhan, Xing Wang

TL;DR

This work tackles camouflaged object detection (COD), where object boundaries are difficult to discern due to background camouflage. It proposes CS$^3$Net, a compact single-branch network that employs a spotlight shifting strategy to synthesize shadow-based supervisory signals, avoiding additional branches. The architecture centers on Projection Aware Attention (PAA) for refining features guided by shadow cues and Extended Neighbor Connection Decoder (ENCD) for high-resolution decoding, with a training objective that couples shadow-map supervision with standard GT supervision. On standard COD benchmarks, CS$^3$Net achieves state-of-the-art accuracy while reducing MACs by $32.13\%$, demonstrating a strong practical balance between performance and efficiency.

Abstract

Camouflaged object detection (COD) remains a challenging task in computer vision. Existing methods often resort to additional branches for edge supervision, incurring substantial computational costs. To address this, we propose the Co-Supervised Spotlight Shifting Network (CS$^3$Net), a compact single-branch framework inspired by how shifting light source exposes camouflage. Our spotlight shifting strategy replaces multi-branch designs by generating supervisory signals that highlight boundary cues. Within CS$^3$Net, a Projection Aware Attention (PAA) module is devised to strengthen feature extraction, while the Extended Neighbor Connection Decoder (ENCD) enhances final predictions. Extensive experiments on public datasets demonstrate that CS$^3$Net not only achieves superior performance, but also reduces Multiply-Accumulate operations (MACs) by 32.13% compared to state-of-the-art COD methods, striking an optimal balance between efficiency and effectiveness.

Shifting Spotlight for Co-supervision: A Simple yet Efficient Single-branch Network to See Through Camouflage

TL;DR

This work tackles camouflaged object detection (COD), where object boundaries are difficult to discern due to background camouflage. It proposes CSNet, a compact single-branch network that employs a spotlight shifting strategy to synthesize shadow-based supervisory signals, avoiding additional branches. The architecture centers on Projection Aware Attention (PAA) for refining features guided by shadow cues and Extended Neighbor Connection Decoder (ENCD) for high-resolution decoding, with a training objective that couples shadow-map supervision with standard GT supervision. On standard COD benchmarks, CSNet achieves state-of-the-art accuracy while reducing MACs by , demonstrating a strong practical balance between performance and efficiency.

Abstract

Camouflaged object detection (COD) remains a challenging task in computer vision. Existing methods often resort to additional branches for edge supervision, incurring substantial computational costs. To address this, we propose the Co-Supervised Spotlight Shifting Network (CSNet), a compact single-branch framework inspired by how shifting light source exposes camouflage. Our spotlight shifting strategy replaces multi-branch designs by generating supervisory signals that highlight boundary cues. Within CSNet, a Projection Aware Attention (PAA) module is devised to strengthen feature extraction, while the Extended Neighbor Connection Decoder (ENCD) enhances final predictions. Extensive experiments on public datasets demonstrate that CSNet not only achieves superior performance, but also reduces Multiply-Accumulate operations (MACs) by 32.13% compared to state-of-the-art COD methods, striking an optimal balance between efficiency and effectiveness.
Paper Structure (11 sections, 7 equations, 5 figures, 2 tables)

This paper contains 11 sections, 7 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. The Proposed Method
  3. Spotlight Shifting Strategy
  4. Projection Aware Attention
  5. Extended Neighbor Connection Decoder
  6. Loss Function and Co-supervision
  7. Experiments And Results
  8. Experiment Settings
  9. Comparison with the State-of-the-Art Methods
  10. Ablation Study
  11. Conclusion

Figures (5)

  • Figure 1: An illustration of spotlight shifting strategy, the shadow projection over the balloon dog changes dramatically as the spotlight shifts from left to right. These changes highlights the object’s contours, making them more distinct and easier to detect.
  • Figure 2: Comparison of model performance ($F^w_\beta$ on NC4K), parameters and MACs across state-of-the-art COD methods.
  • Figure 3: The architecture of the proposed CS$^3$Net. CS$^3$Net operates as a single-branch network utilizing spotlight shifting strategy for co-supervision, it consists of two key modules: the Projection Aware Attention (PAA) and the Extended Neighbor Connection Decoder (ENCD) to integrate knowledge gained from co-supervision.
  • Figure 4: Visual performance of CS$^3$Net and other highly competitive COD methods.
  • Figure 5: Visualization of feature maps before and after proposed modules: Spotlight Shifting Co-supervision and PAA.