Unable to reproduce paper metrics, questions about paper results

[wjaycyf]

Hi, I downloaded the train_blur_bicubic and val_blur_bicubic datasets of type Blur + Low Resolution according to the dataset link you provided，

Next, I run . /preprocessing/generate_reds4.py and . /preprocessing/generate_flow.py, which generated the corresponding reds4 validation set and the corresponding optical flow file as follows，

Using the above files as a test set, gt is the corresponding files taken from the dataset links train_orig_part0.zip and train_orig_part1.zip.
Next, I run python main.py --test --config_path experiment.cfg with the provided pre-training weights and the accuracy results of the test are as follows，



The results are quite different from the results of your paper, may I ask which type of data from REDS did you use for the experiment? If I want to reproduce the results of your paper, which type of dataset do I need to download? Looking forward to your reply.

[GeunhyukYouk]

Hello,
It seems that you have downloaded the wrong GT data.
The GT data should be train_sharp and val_sharp, not train_origin. Please refer to the image below.

Also, generate_flow.py is required only for generating pseudo-GT optical flow used in training, so it is not necessary when testing with the provided pretrained model.