Technical Report for the ICRA 2026 GOOSE 2D Fine-Grained Semantic Segmentation Challenge: Pretraining-Diverse Ensemble of Foundation Vision Encoders for Robust Outdoor Scene Understanding

2026-06-22

AI Summary

This work addresses the challenge of fine-grained semantic segmentation across 56 classes in unstructured outdoor scenes captured by four distinct camera platforms. The authors integrate diverse pretrained vision encoders—including DINOv2, SigLIP, and InternImage—with a Mask2Former decoder, leveraging large crop sizes, extended training schedules, exponential moving average, and multi-scale flip test-time augmentation. They further introduce a novel class-weighted ensemble strategy based on validation-set IoU scores. Systematic analysis reveals that the choice of pretraining strategy exerts a substantially greater influence on performance than model parameter count or decoder architecture. The proposed approach achieves 75.40% mIoU on the official GOOSE test set, securing second place in the competition.

Abstract

This report presents our solution for the ICRA 2026 GOOSE 2D Fine-Grained Semantic Segmentation Challenge, which requires parsing unstructured outdoor scenes from four camera platforms into 56 fine-grained categories. Our approach pairs foundation vision encoders (including DINOv3, SigLIP2, and InternImage) with a Mask2Former decoder, and trains them with a strong recipe including long training schedules, exponential moving average, a larger crop size, and multi-scale plus flip test-time augmentation. The three encoders, chosen for their complementary pretraining objectives, are combined into a pretraining-diverse ensemble through per-class validation-IoU weighting. Evaluated on the official GOOSE test set, our submission achieves 75.40% composite mIoU and wins the second place of the challenge. Our study further shows that the encoder's pretraining recipe, rather than its parameter count or the decoder design, is the dominant factor for accuracy on this benchmark.

Problem

• fine-grained semantic segmentation
• outdoor scene understanding
• multi-camera perception
• 2D semantic parsing

Innovation

• pretraining-diverse ensemble
• foundation vision encoders
• fine-grained semantic segmentation
• outdoor scene understanding
• validation-IoU weighting