DINOv3 ViT-L backbone wins ICRA 2026 GOOSE segmentation challenge

Winning the ICRA 2026 GOOSE challenge required parsing outdoor field-robot imagery into 64 fine-grained classes (11 non-void coarse categories). The first-place team (Park et al.) leveraged a self-supervised DINOv3 ViT-L/16 backbone coupled with a ViT-Adapter and a Mask2Former mask-classification decoder. They introduced a coarse-category auxiliary loss on the global [CLS] token to improve hierarchical understanding. At inference, they employed multi-scale and horizontal-flip test-time augmentation and an ensemble of their top-3 checkpoints (selected via Codabench scores).

The final submission achieved a composite score of 76.57% (69.32% fine-class mIoU, 83.81% category-level mIoU), securing first place on the final leaderboard. The work demonstrates that self-supervised vision transformers combined with modern segmentation decoders and strategic ensembling can robustly handle complex outdoor scenes. The approach is particularly relevant for field robotics applications such as autonomous navigation in unstructured environments.