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 (2505.06991v1)

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.

Summary of "Technical Report for ICRA 2025 GOOSE 2D Semantic Segmentation Challenge"

The paper "Technical Report for ICRA 2025 GOOSE 2D Semantic Segmentation Challenge" presents a sophisticated semantic segmentation framework tailored to effectively parse outdoor scenes into nine distinct semantic categories in real-world environments. This framework was devised by team ACVLAB for the ICRA 2025 GOOSE 2D Semantic Segmentation Challenge, which evaluates models on their capability to generalize across variable outdoor conditions using data from four robotic platforms with differing camera setups and viewpoints.

The key contributions of this paper involve notable advancements in the combination of several state-of-the-art methodologies. The framework's backbone leverages a Swin Transformer, augmented by Rotary Position Embedding (RoPE) to enhance spatial generalization. This combination permits the model to effectively address resolution and scale variations. The incorporation of a Color Shift Estimation-and-Correction module tackles the inherent challenge of illumination inconsistencies typical in diverse and uncontrolled lighting conditions. Additionally, a quantile-based label denoising strategy is employed, specifically targeting high-error pixel data by downweighting the top 2.5% of error-prone pixel labels during training, thus maintaining robustness against annotation noise.

The performance of the proposed framework was evaluated using the GOOSE test set, yielding a competitive mean Intersection over Union (mIoU) of 0.848. This result underscores the framework's efficacy in mitigating common challenges such as domain shifts, annotation imperfections, and lighting variations, which are all prevalent in real-world robotic perception tasks.

From a practical standpoint, this research implies a substantive step forward in equipping robotic systems with reliable scene understanding capabilities, essential for their operation in unpredictable and diverse outdoor settings. The findings suggest that the fusion of transformer-based architectures with robustness-oriented strategies can significantly enhance the fidelity and stability of semantic segmentation outputs.

The approach presented in this paper demonstrates strong numerical results through a well-designed model evaluation. It offers a foundation for future research exploring further enhancements in model robustness and generalization, particularly as it applies to complex outdoor scenarios proliferated by variable lighting and sensor noise. Continued exploration of advanced positional encoding techniques and adaptive error mitigation strategies could potentially lead to even more reliable implementations across various domains of computer vision in robotics.