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
156 stars 20 forks source link

Scaling of point clouds #9

Open thrhgh opened 3 years ago

thrhgh commented 3 years ago

Is there any way to register two point clouds with a different scale with this algorithm?

XiaoshuiHuang commented 3 years ago

The current global feature is invariant to scale difference, so it is difficult to solve the scale difference. If there is a scale-attentive feature available, this framework could be utilized to solve the scale difference by replacing the PointNet with the new scale-attentive feature. Also, the current LK algorithm needs to be replaced with a scale+rotation+translation estimation algorithm.

thrhgh commented 3 years ago

First of all thanks for sharing your projects source code and for doing this amazing work. Also thanks for your helpful reply. So let's assume I would generate a point cloud from a depth camera and want to register a 3D Object (for instance a point cloud of a desk) on this camera depth point cloud. I assume this would not be possible because of the different scaling from the objects right? Are you aware of any deep learning solution that is able to solve such a problem to register a different scaled object on a camera point cloud scene?