D
Data
Text
B2Net: Camouflaged Object Detection via Reused Boundary-Aware Modules
How the B2Net segmentation network improves camouflaged object detection by refusing to commit to an edge prior early. Built on a Pyramid Vision Transformer backbone, it refines each scale with a Residual Feature Enhanced Module, then applies a Boundary Aware Module twice - fusing low-level spatial detail with high-level semantics to sharpen edge cues - and a Cross-scale Boundary Fusion Module that merges boundary and object features top-down. The learner sees why reusing boundary reasoning across stages beats a single early edge prediction, and how B2Net outperforms 15 prior methods on standard COD benchmarks.