Why vespcn is not better than EDSR in same training datasets?

LoSealL / VideoSuperResolution

A collection of state-of-the-art video or single-image super-resolution architectures, reimplemented in tensorflow.

MIT License

1.62k stars 296 forks source link

Why vespcn is not better than EDSR in same training datasets? #35

Closed HymEric closed 5 years ago

HymEric commented 5 years ago

Why vespcn is not better than EDSR in same train datasets? Any advice to get higher PSNR and SSIM?

LoSealL commented 5 years ago

Many reasons, so I am not able to explain details here. Be aware that maximum pixel displacement (the max output of optical flow) in Tensorflow's VESPCN is only 1 in my implementation. Now I suggest following the PyTorch's implementation and extend the max displacement to around 32.

BTW, EDSR has much more parameters than VESPCN here.

HymEric commented 5 years ago

Thank you.