search

XiaoshuiHuang / fmr

This repository is the implementation of our CVPR 2020 work: "Feature-metric Registration: A Fast Semi-supervised Approach for Robust Point Cloud Registration without Correspondences"
MIT License
151 stars 20 forks source link

Why use the vertex of CAD Model Mesh as a point cloud for experiments? #8

Open MaxChanger opened 3 years ago

MaxChanger commented 3 years ago

I can't understand why the author chose to use the vertex of CAD Model Mesh as the point cloud for the experiment, which is very different from the real point cloud distribution. And does not conform to the community’s settings, such as DCP, PRNet and RPMNet, etc.

Welcome to discuss.

MaxChanger commented 3 years ago

Pinging the author @XiaoshuiHuang.

XiaoshuiHuang commented 3 years ago

The experiments followed the most related work PointetLK to use the vertexes for the experimental validation. You are right, the sampled points from the mesh may be more close to real point cloud distribution. However, the performance of this algorithm might be no different by using the sampled point clouds or vertexes. The algorithm aims to estimate the transformation by minimizing the feature difference of two given point sets. The given point sets could be vertexes or sampled point clouds.

MaxChanger commented 3 years ago

But we simply test that under the same data setting of DCP, FMR cannot converge. Take the liberty to ask, have you ever made a real try instead of theoretical inference? @XiaoshuiHuang Thanks.

XiaoshuiHuang commented 3 years ago

I did not follow the data setting of DCP. In the published results, the 7scenes dataset is uniformed sampled on the surface. The training process is converged and the experimental results are good.

MaxChanger commented 3 years ago

Thank you for your quick answer! If anyone else is following the DCP/RPMNet data setting and cannot converge, welcome to discuss with me.

daichang01 commented 2 months ago

How to train with your own point cloud data