Mutual Iterative Refinement Network for Scribble-Supervised Camouflaged Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Camouflaged Object Detection (COD) is difficult due to high visual similarity between objects and backgrounds.
• Traditional COD methods require costly pixel-level annotations.
• Scribble-Supervised COD (SSCOD) offers a more efficient alternative using sparse annotations.

Purpose of the Study:

• To address challenges in SSCOD, including entangled features and scale variation issues.
• To propose a novel network, MIR-Net, for improved camouflaged object detection under weak supervision.
• To enhance the robustness and accuracy of COD models using minimal annotations.

Main Methods:

• Developed the Mutual Iterative Refinement Network (MIR-Net) with a cross-branch mutual refinement mechanism.
• Introduced Background-driven Foreground Feature Enhancement (BFFE) and Foreground-driven Background Feature Enhancement (FBFE) modules.
• Incorporated a Scale-Invariant Consistency (SIC) loss for improved robustness across different object scales.

Main Results:

• MIR-Net achieved state-of-the-art performance on CAMO, COD10K, and NC4K datasets for SSCOD.
• Outperformed fully supervised Convolutional Neural Network (CNN)-based models.
• Demonstrated competitive performance against fully supervised Transformer-based approaches.

Conclusions:

• MIR-Net effectively disentangles and enhances foreground and background features using weak supervision.
• The proposed methods significantly improve COD accuracy and robustness to scale variations.
• MIR-Net shows strong potential for advancing camouflaged object detection in real-world applications with limited annotations.