WeatherSeg: Semi-Supervised Segmentation for Adverse Weather
WeatherSeg, a newly developed semi-supervised segmentation framework, aims to enhance environmental perception for self-driving vehicles in challenging weather conditions while minimizing annotation expenses. This framework utilizes a Dual Teacher-Student Weight-Sharing Model (DTSWSM) for knowledge transfer from images impacted by weather and incorporates a Classifier Weight Updating Attention Mechanism (CWUAM) that modifies classifier weights according to environmental factors. Assessment results indicate that WeatherSeg surpasses baseline models in both accuracy and resilience under clear, rainy, cloudy, and foggy conditions, positioning it as a viable solution for semantic segmentation in all weather scenarios.
Key facts
- WeatherSeg is a semi-supervised segmentation framework for autonomous driving.
- It addresses environmental perception challenges in adverse weather.
- Reduces annotation costs through semi-supervised learning.
- Integrates Dual Teacher-Student Weight-Sharing Model (DTSWSM).
- Uses Classifier Weight Updating Attention Mechanism (CWUAM).
- Outperforms baseline models in clear, rainy, cloudy, and foggy conditions.
- Published on arXiv under Computer Vision and Pattern Recognition.
- Submission history available on arXiv.
Entities
Institutions
- arXiv