Technical Report for ICRA 2026 GOOSE 2D Fine-Grained Semantic Segmentation Challenge: Leveraging DINOv3 for Robust Outdoor Scene Understanding in Field Robotics

2026-06-16

AI Summary

This work addresses fine-grained semantic segmentation in unstructured off-road scenes, encompassing 64 classes grouped into 11 coarse categories. It presents the first integration of a self-supervised DINOv3 ViT-L/16 backbone with a ViT-Adapter and a Mask2Former decoder, augmented by a coarse-category auxiliary supervision loss applied to the [CLS] token. During inference, multi-scale testing and horizontal flipping are employed, followed by an efficient ensemble of three models selected based on Codabench scores. The proposed method achieves state-of-the-art performance on the official benchmark, securing first place in the challenge with an overall score of 76.57%, comprising a fine-grained mIoU of 69.32% and a category-level mIoU of 83.81%, thereby significantly advancing outdoor fine-grained segmentation capabilities.

Abstract

The GOOSE 2D Fine-Grained Semantic Segmentation Challenge at the ICRA 2026 Workshop on Field Robotics evaluates dense semantic segmentation of off-road imagery over a fine-grained taxonomy of 64 classes and 11 evaluated non-void coarse categories. We present the first-place solution to this challenge. Our solution comprises two complementary improvements: (a) a network-level design that combines a self-supervised DINOv3 ViT-L/16 backbone, a ViT-Adapter, and a Mask2Former mask-classification decoder, together with a coarse-category auxiliary loss on the global [CLS] token; and (b) an inference-time aggregation strategy based on multi-scale and horizontal-flip test-time augmentation and an ensemble of the top three checkpoints selected using Codabench scores. Our method achieves an official composite score of 76.57%, consisting of 69.32% fine-class mIoU and 83.81% category-level mIoU, and ranks first on the final phase leaderboard: www.codabench.org/competitions/14257/#/results-tab.

Problem

• fine-grained semantic segmentation
• field robotics
• off-road imagery
• dense prediction
• outdoor scene understanding

Innovation

• DINOv3
• Fine-grained semantic segmentation
• Mask2Former
• Test-time augmentation
• Field robotics