Zero-Shot Camouflaged Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Camouflaged object detection (COD) aims to identify objects visually integrated with their surroundings.
• Current COD methods struggle with detecting unseen object classes, limiting real-world applicability due to data collection and labeling challenges.
• Existing approaches focus solely on seen classes, failing to generalize to novel categories.

Purpose of the Study:

• To propose a novel zero-shot camouflaged object detection (ZSCOD) framework capable of detecting unseen object classes.
• To address the limitations of existing COD methods in handling novel object categories.
• To develop a method that effectively transfers knowledge from seen to unseen classes for improved detection.

Main Methods:

• Introduction of the ZSCOD framework, comprising a Dynamic Graph Searching Network (DGSNet) and a Camouflaged Visual Reasoning Generator (CVRG).
• DGSNet adaptively captures edge details to enhance COD performance.
• CVRG generates pseudo-features for knowledge transfer from seen to unseen classes, utilizing a dynamic graph searching strategy for improved boundary focus.

Main Results:

• The ZSCOD framework successfully detects camouflaged objects from unseen classes.
• Experimental results demonstrate state-of-the-art performance on both seen and unseen classes using public datasets.
• The proposed dynamic graph searching strategy effectively reduces background influence by focusing on object boundaries.

Conclusions:

• The developed ZSCOD framework offers a significant advancement in zero-shot camouflaged object detection.
• The approach effectively bridges the gap between seen and unseen classes, enhancing the practical utility of COD systems.
• The study also introduces the first benchmark for zero-shot COD, facilitating future research in this domain.