Efficient Unpaired Image Dehazing with Cyclic Perceptual-Depth Supervision
This work improves image dehazing for applications like autonomous driving or photography by reducing artifacts in depth-varying regions, though it is incremental over prior unpaired methods.
The paper tackled the problem of unpaired image dehazing, which avoids costly paired data, by addressing performance degradation near depth borders, achieving a PSNR of 20.36 on the NYU Depth V2 dataset with reduced FLOPs.
Image dehazing without paired haze-free images is of immense importance, as acquiring paired images often entails significant cost. However, we observe that previous unpaired image dehazing approaches tend to suffer from performance degradation near depth borders, where depth tends to vary abruptly. Hence, we propose to anneal the depth border degradation in unpaired image dehazing with cyclic perceptual-depth supervision. Coupled with the dual-path feature re-using backbones of the generators and discriminators, our model achieves $\mathbf{20.36}$ Peak Signal-to-Noise Ratio (PSNR) on NYU Depth V2 dataset, significantly outperforming its predecessors with reduced Floating Point Operations (FLOPs).