search

Algolzw / BSRT

Pytorch code for "BSRT: Improving Burst Super-Resolution with Swin Transformer and Flow-Guided Deformable Alignment", CVPRW, 1st place in NTIRE 2022 BurstSR Challenge (real-world track).
MIT License
179 stars 13 forks source link

Question about magnitude in FLOWDCN #14

Open K1NSA opened 1 year ago

K1NSA commented 1 year ago

Dear author, I see you set the value of offset in 10*tanh(offset), what the difference between original offset and [-10,10] offset?

It seems that the magnitude is a super parameter, did you try other value? e.g. 15,20

Algolzw commented 1 year ago

Hi, the scale is just an empirical value, and you can change it according to your input image sizes.

K1NSA commented 1 year ago

Thanks a lot, and I train my work with/without pretrained Spynet weight. But weirdly, the un pretrained is better than the pretrained. Does this mean my data is not suitable with pretrained model?

Algolzw commented 1 year ago

Yes! It happens in many RAW image datasets because SpyNet is only trained on RGB images.

K1NSA commented 1 year ago

Thanks for your answering, I still wonder, as training size [256,256] and test size[1050,1900]. Does the tanh(offset)*10 works for both of them?

Algolzw commented 1 year ago

It depends on your frames' shifts (large or small). In my experience, you could fine-tune only the DCN part to fit the real shifts using a large patch size (such as 512 or 768).