How to Relieve Distribution Shifts in Semantic Segmentation for Off-Road Environments
Ji-Hoon Hwang, Daeyoung Kim, Hyung-Suk Yoon, Dong-Wook Kim, Seung-Woo Seo
AI summary
Problem
Off-road semantic segmentation degrades due to external domain discrepancies and internal sensor corruption, which cause distribution shifts that existing implicit or style-removal methods fail to adequately address.
Approach
The authors introduce ST-Seg, which generates diverse, realistic styles from ImageNet to broaden the source domain and applies a deep texture manifold loss to stabilize local feature representations during training.
Key results
- Explicitly expands source domain style distribution using realistic, unbiased style sampling
- Stabilizes texture features via a deep texture manifold regularization loss
- Maintains robust segmentation performance across both external and internal distribution shifts
- Outperforms existing baselines on diverse off-road datasets with challenging real-world scenarios
Why it matters
Enables reliable autonomous navigation in unstructured off-road environments by ensuring accurate terrain perception despite environmental changes and sensor noise.
Abstract
Semantic segmentation is crucial for autonomous navigation in off-road environments, enabling precise classifica- tion of surroundings to identify traversable regions. However, distinctive factors inherent to off-road conditions, such as source-target domain discrepancies and sensor corruption from rough terrain, can result in distribution shifts that alter the data differently from the trained conditions. This often leads to inaccurate semantic label predictions and subsequent failures in navigation tasks. To address this, we propose ST-Seg, a novel framework that expands the source distribution through style expansion (SE) and texture regularization (TR). Unlike prior methods that implicitly apply generalization within a fixed source distribution, ST-Seg offers an intuitive approach for distribution shift. Specifically, SE broadens domain coverage by generating diverse realistic styles, augmenting the limited style information of the source domain. TR stabilizes local texture representation affected by style-augmented learning through a deep texture manifold. Experiments across various distribution- shifted target domains demonstrate the effectiveness of ST- Seg, with substantial improvements over existing methods. These results highlight the robustness of ST-Seg, enhancing the real-world applicability of semantic segmentation for off-road navigation.