search

YanchaoYang / Conditional-Prior-Networks

This is the implementation of our paper: Conditional Prior Networks for Optical Flow
20 stars 1 forks source link

What dose the Fig3‘s mean？ #2

Open lhanjie opened 5 years ago

lhanjie commented 5 years ago

Thank you very much for your open source code, but I have some question when I read your paper. I don't know what the input of the upper branch refers to. Any optical flow? Or is it groundtruth of optical flow? What is the output of CPN network? Can the optical flow in the picture be obtained by only one color image during the test?

YanchaoYang commented 4 years ago

Hi there, the upper branch takes flow in both training and testing. During training takes the ground-truth flow, however, in testing (when we train an optical flow prediction network that takes in two images and output an OF using Eq. 7) the upper branch of CPN takes the predicted flow, and return a score using Eq.3. So the output of the decoder of CPN is flow, but the output of the CPN is a scalar by Eq.3. As mentioned before, CPN is not making any prediction from a single image, but providing a scalar as a condition prior. But a network trained using Eq. 7 takes care of OF prediction using two images.