Skip to content

Latest commit

 

History

History
44 lines (31 loc) · 1.85 KB

README.md

File metadata and controls

44 lines (31 loc) · 1.85 KB

RCDNet (CVPR'2020)

A Model-Driven Deep Neural Network for Single Image Rain Removal (CVPR'2020) 
@inproceedings{wang2020model,
  title={A model-driven deep neural network for single image rain removal},
  author={Wang, Hong and Xie, Qi and Zhao, Qian and Meng, Deyu},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={3103--3112},
  year={2020}
}

rcdnet


Quantitative Result

The metrics are PSNR/SSIM. Both are evaluated on RGB channels.

Method Rain200L Rain200H Rain800 Rain1200 Rain1400
rcdnet_c32s17n4 39.14/0.986 29.43/0.900 27.75/0.872 32.62/0.917 31.28/0.919

Network Complexity

Input shape Flops Params
(3, 256, 256) 194.54GFlops 2.97M

Help Wanted

The results obtained from test.py deviates from the true value. It is speculated that the network weights has not been saved correctly. The experimental results given above are from the evaluation process of the last epoch (by checking log files). The evaluation process, which is called from EvalHook, differs from the testing process of test.py in that it loads the network weights in memory directly, without going through the serialization process.

  • Updated on 2022.4.6: Seems that it's not the issue of saving weights. Quite confused now.