Revealing Hidden Context in Camouflage Instance Segmentation

Thanh-Hai Phung, Hong-Han Shuai; Proceedings of the Asian Conference on Computer Vision (ACCV), 2024, pp. 2319-2336

Predicting the instance-level masks of objects hidden in complex contexts is the goal of Camouflage Instance Segmentation (CIS), a task complicated by the striking similarities between camouflaged objects and their backgrounds. This challenge is further heightened by the diverse appearances of camouflage objects, including varying angles, partial visibilities, and ambiguous morphologies. Prior works considered classifying pixels in a high uncertainty area without considering their contextual semantics, leading to numerous false positives. We proposed a novel method called Mask2Camouflage, which simultaneously enhances the modeling of contextual features and refines instance-level predicted maps. Mask2Camouflage leverages multi-scale features to integrate the extracted features from the backbone. Then, a Global Refinement Cross-Attention Module (GCA) is introduced to complement the foreground mask and background mask each other to reduce the false positive. Furthermore, by simulating a global shift clustering process, we present the Global-Shift Multi-Head Self-Attention (GSA), which enables the object query to capture not only information from earlier features but also their structural concepts, thereby reducing intra-class issues in the camouflage object detection task when validated with evaluated data. Compared with 15 state-of-the-art approaches, our Mask2Camouflage significantly improves the performance of camouflage instance segmentation.