The final test stage tests the calculation time and what is the input size of flops. Since the standard 4K is 3840x2160, and the crop-size in the program you provided is [1080, 2040], [1020, 540]. I'm confused about this, can you clarify the size of the input?
Posted by: PixelBE @ Feb. 21, 2023, 11:52 a.m.follow. for 2x I guess it should be 1920x1080
Posted by: tuvovan @ Feb. 22, 2023, 5:33 a.m.I just checked the image in the validation input, it seems like the size is (540, 1020, 3), which will be 1020, 2040,3 after going through the model? is it 4K?
Posted by: tuvovan @ Feb. 22, 2023, 5:56 a.m.In https://github.com/eduardzamfir/NTIRE23-RTSR/blob/master/demo/runtime_demo.py lines 69 and 123 indicate that the size of the calculation time is [1020,540], and the task of the competition is 1080p to 4k, so I am confused.
Posted by: PixelBE @ Feb. 22, 2023, 6:03 a.m.Hello, thanks for your question / bug detected.
during the test phase, the output image will be 4K (3840, 2160).
The input of the network will be (1920, 1080) in the X2 track, and (1280, 720) in the X3 track, the LR images are generated as indicated in our repository.
The HR images are native 4k/8K resolution.
The current validation test set includes images from DIV2K and an internal dataset of 4K and 8K images.
Only for validation purposes we used [1020, 540] as input, since DIV2K has no 4K images.
We are updating the repo so when you measure runtime, is always considering that the output will be 4K.
https://github.com/eduardzamfir/NTIRE23-RTSR/blob/master/demo/runtime_demo.py
- Organizers.
Posted by: nanashi @ Feb. 22, 2023, 10:33 a.m.