Camouflaged instance segmentation (CIS) focuses on handling instances that attempt to blend into the background. However, existing CIS methods emphasize global interactions but overlook hidden clues at various scales, resulting in inaccurate recognition of camouflaged instances. To address this, we propose a multi-scale pooling network (MSPNet) to mine the hidden cues offered by the camouflaged instances at various scales. The network achieves an enhanced fusion of multi-scale information mainly through multilayer pooling. Specifically, the pyramid pooling transformer (P2T) is utilized as a ro...
