RCVN-For-Video-Enhancement
NTIRE 2021 video/multi-frame challenges
Quality enhancement of heavily compressed videos: Track 1 Fixed QP, Fidelity
The test results of Track 1:
| rank | PSNR | rank | PSNR |
|---|---|---|---|
| 1 | 32.52 | 7 | 31.75 |
| 2 | 32.49 | 8 | 31.65 |
| 3 | 32.04 | 9 | 31.62 |
| 4 | 31.90 | Ours | 31.59 |
| 5 | 31.86 | 11 | 31.37 |
| 6 | 31.78 | 12 | 31.13 |
We implemented our methods with
- python 3.8.8
- pytorch 1.8.0
- cuda 11.1
- cudnn 8.0.5.
In our machine, it cost about 13.72 seconds to generation a png.
For test, you can follow the steps below.
1. First set full compressed pngs in
/data_path/val_video_png/test_fixed-QP_png/001/xxx.png
/data_path/val_video_png/test_fixed-QP_png/002/xxx.png
...
/data_path/val_video_png/test_fixed-QP_png/010/xxx.png
We have set 10 pngs in the 001 folder for an example.
2. Run the following code to generation enhanced pngs:
python main.py --model rcvn_c --video_numbers 10 --save_results --frame 53. The full enhanced pngs(or 10 enhanced example pngs) results will be generated in
/RCVN_C/test_results/001/xxx.png
...
...