search

FangjinhuaWang / PatchmatchNet

Official code of PatchmatchNet (CVPR 2021 Oral)
MIT License
503 stars 70 forks source link

Question about Adaptive propagation. #27

Closed KyuminHwang closed 3 years ago

KyuminHwang commented 3 years ago

Hi, First of all thanks for your great efforts on this research.

I have a question about adaptive propagation. In model/patchmatch.py at line from 79 to 82.

propogate_depth_sample = F.grid_sample(depth_sample[:, num_depth // 2,:,:].unsqueeze(1), grid, mode='bilinear', padding_mode='border')

Is there have any intuition or knowledge about _num_depth//2_ ? I guessed that num//depth2 is the center of hypothesis planes that is equivalent to zero depth.
I cannot understand this part, clearly. Could you explain for me ?

Thank you for sharing the research results !

FangjinhuaWang commented 3 years ago

Hi,

For each pixel, the input of propagation is the set of hypotheses that are sampled around the estimation d from previous iteration. The index "num_depth//2" corresponds to this estimation d, the other indices correspond to the new samples around d.