NTIRE 2022 Stereo Image Super-Resolution Challenge Forum

Go back to competition Back to thread list Post in this thread

> Pretrained models

Dear organizers,

I have two questions regarding the competition rules:
1. Is it allowed to use pretrained optical flow models as a submodule in our solution?
2. How are you going to trace submitted solutions that had used external datasets, pretrained models on bicubic degradation during the training stage?

Sincerely,
IPG

Posted by: ipg_mrc @ March 14, 2022, 11:05 a.m.

Hi IPG,

(1) Optical flow (or disparity estimation) models pre-trained on external datasets are not allowed. One can use traditional approaches or train these models on the provided training dataset in an unsupervised manner.
(2) We will re-train the networks using the codes submitted by the participants and exclude results with PSNR difference larger than a threshold.

Best Regards,
Longguang

Posted by: Rogen @ March 15, 2022, 3:29 a.m.

Thank you for your answer. Regarding the second issue, does it mean that the final results for this challenge will take several weeks (until organizers will check for the reproducibility)?

Posted by: ipg_mrc @ March 15, 2022, 10:42 a.m.
Post in this thread