Technical Report for ICRA 2025 GOOSE 2D Semantic Segmentation Challenge: Leveraging Color Shift Correction, RoPE-Swin Backbone, and Quantile-based Label Denoising Strategy for Robust Outdoor Scene Understanding

TL;DR

This work tackles robust outdoor 2D semantic segmentation under cross-domain and annotation-noise conditions by building a MaskDINO-based pipeline enhanced with three robustness modules: Color Shift Estimation-and-Correction (CSEC) to normalize illumination, a RoPE-ViT backbone to improve spatial generalization across scales, and a quantile-based denoising strategy to downweight unreliable labels. The CSEC component integrates a Color Shift Estimation (COSE) step and a Color Modulation (COMO) step to stabilize color and brightness, while RoPE-ViT introduces Rotary Positional Embedding into self-attention to better capture spatial relationships. Quantile-based denoising removes the top 2.5% of high-error pixels via a percentile-based data filter, improving training stability. On the GOOSE dataset, the approach achieves a test mIoU of 84.8%, with ablations showing significant gains from both RoPE and CSEC, and further gains when denoising is applied, highlighting the value of combining color correction, advanced positional encoding, and error-aware training for real-world robotic perception.

Abstract

This report presents our semantic segmentation framework developed by team ACVLAB for the ICRA 2025 GOOSE 2D Semantic Segmentation Challenge, which focuses on parsing outdoor scenes into nine semantic categories under real-world conditions. Our method integrates a Swin Transformer backbone enhanced with Rotary Position Embedding (RoPE) for improved spatial generalization, alongside a Color Shift Estimation-and-Correction module designed to compensate for illumination inconsistencies in natural environments. To further improve training stability, we adopt a quantile-based denoising strategy that downweights the top 2.5\% of highest-error pixels, treating them as noise and suppressing their influence during optimization. Evaluated on the official GOOSE test set, our approach achieved a mean Intersection over Union (mIoU) of 0.848, demonstrating the effectiveness of combining color correction, positional encoding, and error-aware denoising in robust semantic segmentation.

Figures (3)

• Figure 1: Architecture of Robust Outdoor Scene Understanding with Color Shift Correction, RoPE-Swin Backbone, and Quantile-based Denoising
• Figure 2: Comparison of segmentation results between CSEC-enhanced model and baseline
• Figure 3: Comparison of segmentation results with original and RoPE backbone