Dual-Path Transformer with Spatial-Channel Attention for Enhanced Panoptic Segmentation

Panoptic segmentation is a key task in computer vision, aiming to achieve both instance-level and semantic-level segmentation. However, existing methods generally suffer from problems such as insufficient feature utilization, limited instance differentiation, and insufficient multi-scale information integration. To address these challenges, this paper proposes a novel panoptic segmentation architecture with two key innovations: (1) a feature interaction module based on a spatial-channel dual-dimensional attention mechanism, which significantly enhances the contextual information capture and multi-scale feature representation capabilities by accurately modeling the dependencies between different spatial locations and channel dimensions; (2) an innovative dual-path Transformer decoder, which effectively enhances instance boundary differentiation and global information integration capabilities through the parallel optimization design of the pixel-query interaction path and the query self-update path. In the standard panoptic segmentation benchmark, our method achieves superior performance on the MS-COCO and Cityscapes datasets, reaching 63.1% and 52.8% respectively, providing a new solution for high-quality panoptic segmentation in practical applications.