WeatherProof Dataset
收藏资源简介:
WeatherProof Dataset是由加州大学洛杉矶分校创建的第一个具有精确配对清晰和恶劣天气图像对的语义分割数据集,包含超过174,000张图像。该数据集通过确保清晰和恶劣天气图像之间的底层语义标签相同,提供了一个受控的测试平台,其中性能退化主要归因于天气因素。该数据集不仅支持新的训练范式,还改进了清晰和退化分割之间的性能评估差距。通过使用这种配对数据集进行训练,可以分离学习新场景和学习对天气影响的恢复力,从而提高模型在恶劣天气场景下的性能。
The WeatherProof Dataset, created by the University of California, Los Angeles (UCLA), is the first semantic segmentation dataset with precisely paired clear and adverse weather image pairs, containing over 174,000 images. By ensuring that the underlying semantic labels are identical across both clear and adverse weather images, this dataset provides a controlled testbed where performance degradation is primarily attributed to weather-related factors. This dataset not only supports novel training paradigms but also enhances the validity of performance comparisons between clear-weather and degraded semantic segmentation tasks. Training with this paired dataset allows researchers to disentangle the learning of novel scenes and the learning of resilience against weather perturbations, thereby improving model performance in adverse weather scenarios.




