search

qinzheng93 / GraphSCNet

[CVPR2023] Deep Graph-based Spatial Consistency for Robust Non-rigid Point Cloud Registration
MIT License
78 stars 4 forks source link

Warp function for points in the no overlap region of the src point cloud ? #1

Closed 2019EPWL closed 1 year ago

2019EPWL commented 1 year ago

Hi, I have a question about the teaser figure in the GraphSCNet paper. The warp function (10) is defined based on the predicted inlier correspondences. So, that means the paper doesn't deal with the outlier points, but the teaser and results figures show the dense shape reconstruction mesh, I am confused about that.

qinzheng93 commented 1 year ago

Yes, it is very difficult to accurately infer the motion of the outlier points as there are no inlier correspondences within them. So the motion of these points is mainly constraint (or guessed, which I prefer) by the ARAP term in the N-ICP algorithm. The problem is more severe in low-overlap scenes.

BTW, the teaser shows point clouds, not meshes.

2019EPWL commented 1 year ago

So the EPE, ACCS, and ACCR in the result table are only calculated on the predicted inliers, not on the all src point cloud.

qinzheng93 commented 1 year ago

No, following NDP, the metrics are computed on ALL points.