Recurrent Squeeze-and-Excitation Context Aggregation Net for Single Image Deraining (ECCV'2018)
@inproceedings{li2018recurrent,
title={Recurrent squeeze-and-excitation context aggregation net for single image deraining},
author={Li, Xia and Wu, Jianlong and Lin, Zhouchen and Liu, Hong and Zha, Hongbin},
booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
pages={254--269},
year={2018}
}
Quantitative Result
The metrics are PSNR/SSIM. Both are evaluated on RGB channels.
NOTE: Following the authors' setup, random seed is set to 66 in all experiments.
| Method | Rain200L | Rain200H | Rain800 | Rain1200 | Rain1400 |
|---|---|---|---|---|---|
| RESCAN(ConvRNN+Add) | 36.09/0.973 | 25.56/0.806 | 26.17/0.835 | 31.91/0.897 | 30.70/0.905 |
| RESCAN(ConvRNN+Full) | 36.07/0.973 | 25.88/0.810 | 25.89/0.826 | 31.93/0.896 | 30.75/0.904 |
| RESCAN(ConvGRU+Add) | 36.74/0.977 | 26.31/0.828 | 26.55/0.843 | 32.25/0.905 | 30.88/0.910 |
| RESCAN(ConvGRU+Full) | 36.92/0.978 | 26.55/0.834 | 26.70/0.840 | 32.24/0.904 | 30.92/0.910 |
| RESCAN(ConvLSTM+Add) | 36.83/0.977 | 26.34/0.829 | 26.69/0.847 | 32.18/0.903 | 30.92/0.909 |
| RESCAN(ConvLSTM+Full) | 36.95/0.978 | 26.49/0.833 | 26.76/0.841 | 32.23/0.904 | 31.02/0.911 |
Pretrained models can be downloaded from here
Network Complexity
| Method | Input shape | Flops | Params |
|---|---|---|---|
| RESCAN(ConvGRU) | (3, 256, 256) | 32.47GFlops | 150.22k |
| RESCAN(ConvLSTM) | (3, 256, 256) | 43.23GFlops | 197.76k |
| RESCAN(ConvRNN) | (3, 256, 256) | 12.37GFlops | 55.13k |