PRSeg: A Lightweight Patch Rotate MLP Decoder for Semantic Segmentation

The lightweight MLP-based decoder has become increasingly promising for semantic segmentation. However, the channel-wise MLP cannot expand receptive fields, lacking context modeling capacity, which is critical to In this paper, we propose a parametric-free patch rotate operation reorganize pixels spatially. It first divides feature map into multiple groups and then rotates patches within each group. Based on proposed operation, design novel segmentation network, named PRSeg, includes an off-the-shelf backbone Patch Rotate containing Dynamic Blocks (DPR-Blocks). DPR-Block, fully connected layer performed following Module (PRM) exchange spatial information between pixels. Specifically, in PRM, split reserved part rotated along channel dimension according predicted probability of Channel Selection (DCSM), our only part. Extensive experiments ADE20K, Cityscapes COCO-Stuff 10K datasets prove effectiveness approach. We expect that PRSeg can promote development