I encountered severe artifact when testing on VID4 with pretrained model

YapengTian / TDAN-VSR-CVPR-2020

TDAN: Temporally-Deformable Alignment Network for Video Super-Resolution, CVPR 2020

MIT License

400 stars 62 forks source link

I encountered severe artifact when testing on VID4 with pretrained model #41

Closed T5eng closed 3 years ago

T5eng commented 3 years ago

as described in the title, here is some cases. But case calendar is alright, same as visual quality in the paper.

It would be great if you could help solve this problem.

YapengTian commented 3 years ago

The released model is for BI setting. It means that you need to use Matlab resize function with the "bicubic" option to downsample HR frames to generate LR frames.

T5eng commented 3 years ago

In my experiment, LR frames are generated with cv2.resize(... , interpolation=cv2.INTER_CUBIC) and frames in calendar set are all looking good. Artifact exists only in foliage, city and walk set... so weird...

YapengTian commented 3 years ago

The functions in OpenCV and Matlab should be different. I guess you will have the similar issue if you test the generated LR frames using other BI models. But, I also cannot understand why the results of the calendar sequence are good.

T5eng commented 3 years ago

I tried to generate LR with Matlab bicubic interpolation. But artifact still exists.....

YapengTian commented 3 years ago

Could you use this code: https://github.com/yulunzhang/RDN/blob/master/RDN_IR/RDN_TestCode/RDN_SR_RGB/Prepare_TestData_HR_LR.m? The BI setting.

T5eng commented 3 years ago

It works. It seems this and some other similar methods are functioning on very specifically degredation....