Skip to content

Latest commit

 

History

History
34 lines (24 loc) · 2.41 KB

QA.md

File metadata and controls

34 lines (24 loc) · 2.41 KB
Raw

Frequently Asked Questions

1. How to reproduce your results in the PIRM18-SR Challenge (with low perceptual index)?

First, the released ESRGAN model in the GitHub (RRDB_ESRGAN_x4.pth) is different from the model we submitted in the competition. We found that the lower perceptual index does not always guarantee a better visual quality. The aims for the competition and our ESRGAN work will be a bit different. We think the aim for the competition is the lower perceptual index and the aim for our ESRGAN work is the better visual quality.

More analyses can be found in Sec 4.1 and Sec 5 in PIRM18-SR Chanllenge report. It points out that PI (perceptual index) is well correlated with the human-opinion-scores on a coarse scale, but it is not always well-correlated with these scores on a finer scale. This highlights the urgent need for better perceptual quality metrics.)

Therefore, in the PIRM18-SR Challenge competition, we used several tricks for the best perceptual index (see Section 4.5 in the paper).

Here, we provid the models and codes used in the competition, which is able to produce the results on the PIRM test dataset (we use MATLAB 2016b/2017a):

Group Perceptual index RMSE
SuperSR 1.978 15.30
  1. Download the model and codes from GoogleDrive
  2. Put LR input images in the LR folder
  3. Run python test.py
  4. Run main_reverse_filter.m in MATLAB as a post processing
  5. The results on my computer are: Perceptual index: 1.9777 and RMSE: 15.304

2. How do you get the perceptual index in your ESRGAN paper?

In our paper, we provide the perceptual index in two places.

1). In the Fig. 2, the perceptual index on PIRM self validation dataset is obtained with the model we submitted in the competition. Since the pupose of this figure is to show the perception-distortion plane. And we also use the post-precessing here same as in the competition.

2). In the Fig.7, the perceptual indexs are provided as references and they are tested on the data generated by the released ESRGAN model RRDB_ESRGAN_x4.pth in the GiuHub. Also, there is no post-processing when testing the ESRGAN model for better visual quality.