Improved Techniques for Learning to Dehaze and Beyond: A Collective Study
share
This paper reviews the collective endeavors by the team of authors in exploring two interlinked important tasks, based on the recently released REalistic Single Image DEhazing (RESIDE) benchmark: i) single image dehazing as a low-level image restoration problem; ii) high-level visual understanding (e.g., object detection) from hazy images. For the first task, the authors investigated a variety of loss functions, and found the perception-driven loss to improve dehazing performance very notably. For the second task, the authors came up with multiple solutions including using more advanced modules in the dehazing-detection cascade, as well as domain-adaptive object detectors. In both tasks, our proposed solutions are verified to significantly advance the state-of-the-art performance. The GitHub repository URL is: https://github.com/guanlongzhao/dehaze
READ FULL TEXT
