GOOSE-M2F: Adapting Mask2Former for High-Fidelity, Long-Tailed Fine-Grained Semantic Segmentation in Unstructured Outdoor Terrain
Researchers have developed GOOSE-M2F, a specialized version of Mask2Former designed for the GOOSE 2D Fine-Grained Semantic Segmentation (FGSS) Challenge. This new model addresses the challenge of segmenting 64 fine-grained classes in unstructured outdoor terrain, particularly focusing on rare classes with very few pixels. GOOSE-M2F incorporates several enhancements, including an increased number of object queries, a feature refinement module with attention mechanisms, and an auxiliary supervision head to improve gradients for rare classes. The model achieved third place in the challenge with a composite mIoU of 70.08%. AI
IMPACT Enhances fine-grained semantic segmentation capabilities for rare classes in complex outdoor environments.