search

Gorilla-Lab-SCUT / LPDC-Net

CVPR2021 paper "Learning Parallel Dense Correspondence from Spatio-Temporal Descriptorsfor Efficient and Robust 4D Reconstruction"
MIT License
29 stars 1 forks source link

Some questions about the article #1

Closed haolyuan closed 3 years ago

haolyuan commented 3 years ago

Thanks for your releasing! I have some questions about the part 'Cross-time Correspondence Modeling' in your article. I am confused about the associate position. Given a point p0 in initial time, how can I get the associate position pk in other time? Is the points among different time corresponding?

tangjiapeng commented 3 years ago

For the point sampled from dynamic surface sequence, we can know the dense correspondences between the points among different time. Please refer to the export_pointcloud function in the scripts/sampling_mesh.py. However, if the point p0 is not lying in the suface in time t0, we don't know the ground truth of associate position pk in other time. We just constrain that the occupancy prediction of pk should be consistent with the occupancy label of p0. In other word, the occupancy values of points are always invariant along the temporal evolution.

haolyuan commented 3 years ago

Thank you for your explanation! I know that the values of f(p0) in initial time and f(pk) in k time should be consistent in part ' Cross-time Correspondence Modeling'. However, given a point p0 in initial time, is the associate position pk in other time known in data set? In other words, is the points in the same body part sampled in different point clouds corresponding previously in data set to train the transform process g?

tangjiapeng commented 3 years ago

Yes! According to the ground-truth surface mesh sequences, we can sample some point cloud trajectories.

haolyuan commented 3 years ago

Thanks for your explanation!

tangjiapeng commented 3 years ago

You are welcome! Hope I can help you.