Point Discriminative Learning for Data-efficient 3D Point Cloud Analysis

TL;DR

This paper introduces PointDisc, a self-supervised learning method for 3D point cloud analysis that improves data efficiency by enforcing local and global feature consistency, reducing the need for large labeled datasets.

Contribution

Proposes PointDisc, a novel point discriminative learning approach that enhances data-efficient 3D point cloud classification and segmentation through a new loss on intermediate features.

Findings

Improves classification and segmentation accuracy with less labeled data.

Learns features that effectively capture local and global geometry.

Demonstrates robustness across multiple 3D tasks.

Abstract

3D point cloud analysis has drawn a lot of research attention due to its wide applications. However, collecting massive labelled 3D point cloud data is both time-consuming and labor-intensive. This calls for data-efficient learning methods. In this work we propose PointDisc, a point discriminative learning method to leverage self-supervisions for data-efficient 3D point cloud classification and segmentation. PointDisc imposes a novel point discrimination loss on the middle and global level features produced by the backbone network. This point discrimination loss enforces learned features to be consistent with points belonging to the corresponding local shape region and inconsistent with randomly sampled noisy points. We conduct extensive experiments on 3D object classification, 3D semantic and part segmentation, showing the benefits of PointDisc for data-efficient learning. Detailed analysis demonstrate that PointDisc learns unsupervised features that well capture local and global geometry.