Giters

LoSealL / VideoSuperResolution

A collection of state-of-the-art video or single-image super-resolution architectures, reimplemented in tensorflow.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Issue with Eval of VESPCN in VSRTorch.

AIRedWood opened this issue · comments

commented

Hello, I trained vespcn with VSRtorch and tested it, but the results of the test are not ideal, psnr is only 23.50. I have considered several reasons:

  1. There is no test only on the Y channel (this problem occurs when I use VSR for testing, do not use --l_only, psnr=23., use --l_only, psnr=25.) I view The parameters of VSRtorch found --output_color. But when I use --output_color RGB and --output_color GRAY , psnr doesn't change at all, it's 23.50. Excuse me, where is it wrong?
  2. When I use the vid4 dataset to test it in the VSR. The progress shows 1, 2, ..., 163/163. But in VSRtorch. The progress shows 1, 2, ..., 171/171. (vid4 is four groups of videos, a total of 171 pictures, considering taking the first frame and the last frame, the picture participating in eval should be 163), I would like to ask if this is a display problem or indeed VSRTORch does not consider the interframe Align
commented

1 The psnr print after testing is the same, because it get psnr on directly output and the model always output 1 channel image.
When you choose output RGB, the other 2 channel is bicubical resized to target resolution.

PSNR on luminance channel is always better and because of historical reason, we always compare on that channel.

  1. Maybe I forget to pad head and tail frames in VSR, I will check and fix that.
commented

I have tested with the latest commit:
python run.py --model vespcn --test vid4 and python run.py --model vespcn --infer /data/vid4/input all show progress 1..171/171.