CVPR-2022 PointCLIP:Point Cloud Understanding by CLIP

论文地址:PointCLIP:Point Cloud Understanding by CLIP

论文实现:https://github.com/ZrrSkywalker/PointCLIP

PointCLIP:3D用CLIP预训练编码

Abstract

将CLIP学习到的2D表征迁移到3D领域来,在CLIP编码的点云和3D类别文本之间进行对齐

Method

avatar

关键就是找一个桥梁把3D和2D连接起来就行,把3D点云做了投射到2D平面上变成深度图,这个图像丢给CLIP视觉编码器得到表征

avatar

迁移到3D领域时融合的领域知识的trick

Experiments

avataravatar
avatar
avataravatar