search

lmb-freiburg / flownet2

FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks
https://lmb.informatik.uni-freiburg.de/Publications/2017/IMKDB17/
Other
1k stars 318 forks source link

regarding padding and stride in 1d correlation #125

Closed stalin18 closed 6 years ago

stalin18 commented 6 years ago

Hi, first of all thanks for sharing this great work! I was porting you 1D correlation API in TensorFlow and I've several doubts in the code:

  1. I am wondering why pad size is not equal to max_displacement (in the code, padded width = 2pad_size)? int paddedbottomwidth = bottom[0]->width()+2padsize;

In the paper, it is written that a displacement of 40 is used with stride2 = 1, therefore, isn't padding = 40 required? What pad size do you supply?

  1. single_direction variable: I hope the value -1 corresponds to the case of estimating disparity for left image and +1 corresponds to the case of estimating disparity for right image.

Thank you in advance!

nikolausmayer commented 6 years ago
  1. I'm not sure, but the factor 2 might be a remainder of a network version that didn't do single-direction correlation. In that case we'd pad both directions by max_displacement.
  2. I hope so too. At least that's what it should correspond to.
nikolausmayer commented 6 years ago

(closed due to inactivity)