Skip to content
/ SGP Public

[CVPR 2021 Oral] Self-supervised Geometric Perception

License

Notifications You must be signed in to change notification settings

theNded/SGP

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

20 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SGP: Self-supervised Geometric Perception

[CVPR 2021 Oral] Self-supervised Geometric Perception https://arxiv.org/abs/2103.03114

Introduction

In short, SGP is, to the best of our knowledge, the first general framework for feature learning in geometric perception without any supervision from ground-truth geometric labels.

SGP runs in an EM fashion. It iteratively performs robust estimation of the geometric models to generate pseudo-labels, and feature learning under the supervision of the noisy pseudo-labels.

overview

We applied SGP to camera pose estimation and point cloud registration, demonstrating performance that is on par or even superior to supervised oracles in large-scale real datasets.

Camera pose estimation

Deep image features like CAPS can be trained with relative pose labels generated by 5pt-RANSAC, bootstraped with the handcrafted SIFT feature. They can be later used in robust relative camera pose estimation.

Point cloud registration

Deep 3D features like FCGF can be trained with relative pose labels generated by 3pt-RANSAC, bootstraped by the handcrafted FPFH feature. They can be later used in robust point cloud registration.

Code

Please see code/ for detailed intructions about how to use the code base.

Citation

@inproceedings{yang2021sgp,
  title={Self-supervised Geometric Perception},
  author={Yang, Heng and Dong, Wei and Carlone, Luca and Koltun, Vladlen},
  booktitle={CVPR},
  year={2021}
}

About

[CVPR 2021 Oral] Self-supervised Geometric Perception

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages