CVPR-2022 PointCLIP：Point Cloud Understanding by CLIP

Paper: PointCLIP：Point Cloud Understanding by CLIP

Code: https://github.com/ZrrSkywalker/PointCLIP

PointCLIP: CLIP pre-trained encoding for 3D

Abstract

Transfer representations learned by CLIP in 2D to the 3D domain by aligning CLIP-encoded point clouds with text descriptions of 3D object categories.

Method

The core idea is to bridge 3D and 2D: project a 3D point cloud onto a 2D plane to form a depth map, then pass the resulting image through CLIP’s visual encoder to obtain representations.

Tricks for injecting domain knowledge when adapting to 3D.

Experiments

<HR align=left color=#987cb9 SIZE=1>