How did you construct no_dense version model?

NVlabs / PWC-Net

PWC-Net: CNNs for Optical Flow Using Pyramid, Warping, and Cost Volume, CVPR 2018 (Oral)

Other

1.63k stars 357 forks source link

Open shimhyeonwoo opened 4 years ago

shimhyeonwoo commented 4 years ago

How did you construct no dense version PWC-Net In the PWC-net paper.

I guess you get rid of just dense connection. like,

(corr) -> (128 Ch conv) -> (128 Ch conv) -> (96 Ch conv) -> (64 Ch conv) -> (32 Ch conv) -> (predict flow)

does it right?

shimhyeonwoo commented 4 years ago

if it is the right way of building no_dense version PWC-Net, when you upsample feature map at each level, did you upsampled only the last conv4_4 (32 channel) feature maps to get upsampled feature map(2 channel)?

and what was the input of context network?? was it just conv2_4 (32 channel)?

thank you : )