search

Div99 / W-Stereo-Disp

(NeurIPS 2020 Spotlight) Wasserstein Distances for Stereo Disparity Estimation
https://div99.github.io/W-Stereo-Disp/
MIT License
101 stars 15 forks source link

Question about multi-modal GT #4

Closed Nicholasli1995 closed 3 years ago

Nicholasli1995 commented 3 years ago

Hi, thank you for sharing the code and I have a few questions:

  1. Where can we find the generation of multi-modal ground truth (MM GT) as descripbed in the paper?
  2. Is there an implementation of Wasserstein Loss with MM GT as described by Eq. 12 in the paper?
  3. Did DSGN+CDN (the best 3D AP on KITTI) use MM GT or not? I'm using DSGN and trying to reproduce DSGN+CDN+Wasserstein Loss.
Div99 commented 3 years ago

Hi, thanks for the interest in our paper:

We haven't released the MM GT version yet but plan to do it soon. For DSGN+CDN, we don't use MM GT as we find that that MM GT helps improve training earlier but results in similar final accuracies.

To reproduce results for DSGN, it is easiest to compare CDN-PSMNet to original PSMNet model and apply the changes to DSGN.