XYZ9843/GOOSE-M2F-Segmentation

GOOSE-M2F (Swin-Large)

This model is a task-specific adaptation of Mask2Former (Swin-Large) for the ICRA 2026 GOOSE 2D Semantic Segmentation Challenge in Vienna, Austria.

Performance

• Official Composite mIoU: 70.08% (Ranked at the top of the leaderboard)

Architecture Modifications

1. 200 Object Queries (expanded from 100) for high-density 64-class terrain parsing.
2. Feature Refinement Module (FRM): ASPP-lite + CBAM dual attention to resolve Vegetation and Terrain over-segmentation.
3. Auxiliary Supervision Head: Direct H/4 resolution supervision to bypass query bottlenecks for thin/rare objects.