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

about the spatial scale in L3layer and SPynet #13

Closed K1NSA closed 1 year ago

K1NSA commented 1 year ago

Dear author, Thanks for your work. My work also use a kind of PCD, in the l3 layer ,the H,W =[32,32]. And the SPyNet down sample 32 times , so the input is H,W=[1,1]. Is this possible to use spy net to calculate optical flow?

Algolzw commented 1 year ago

Hi! As my experience, if your input size is different with the required size of SpyNet, you can resize the input to match the SpyNet!

K1NSA commented 1 year ago

Thanks for your answer, by the way, what's the input size of your training data? Is it H,W=[256,256]?

Algolzw commented 1 year ago

For synthetic data, yes its input size is [256 x 256], but for real-world data, its input size is [160 x 160]. But any size should be fine once it can be processed by SpyNet ($> 2^5$).