xvjiarui/GCNet

Introduces the Global Context (GC) block, a lightweight attention mechanism that efficiently models long-range dependencies by computing spatial attention with reduced computational cost through dimensionality reduction (controlled by ratio parameter r). Implemented as a modular operator compatible with mmdetection's Mask R-CNN and other detection/segmentation frameworks, achieving consistent AP improvements (1-2 points on COCO) over baseline detectors with minimal inference overhead.

1,220 stars. No commits in the last 6 months.