HazeBench: A Novel Dataset for Image and Video Dehazing in Natural Environments

Hazy conditions make computer vision tasks difficult, especially when working with real-world videos. Although many video dehazing algorithms have been proposed, their progress is limited by the lack of large, real-world hazy video datasets. To fill this gap, we introduce HazeBench, a dataset compiled from real-world footage captured under various environmental conditions. Unlike many existing datasets, HazeBench does not use ground-truth clear videos, making it more representative of real haze situations. The dataset includes 153 videos (1-15 seconds each) and 65,078 images, grouped into five scene categories: Indoor, Mountains, Night, Road, and Rural Areas. We describe how the dataset was collected, highlight its main features, and show its usefulness through benchmark experiments with video dehazing algorithms. HazeBench provides a valuable resource for developing and testing dehazing methods and supports more reliable computer vision applications in real-world environments.