Self-Supervised Feature Learning from Partial Point Clouds via Pose Disentanglement

TL;DR

This paper introduces a self-supervised learning framework for partial point clouds that disentangles content and pose factors, improving feature representations for various downstream tasks.

Contribution

It proposes a novel framework combining completion, pose regression, and reconstruction networks to enhance feature learning from partial point clouds.

Findings

Outperforms existing self-supervised methods in classification, segmentation, and registration.

Demonstrates better generalization across synthetic and real-world datasets.

Learns robust features by disentangling content and pose factors.

Abstract

Self-supervised learning on point clouds has gained a lot of attention recently, since it addresses the label-efficiency and domain-gap problems on point cloud tasks. In this paper, we propose a novel self-supervised framework to learn informative representations from partial point clouds. We leverage partial point clouds scanned by LiDAR that contain both content and pose attributes, and we show that disentangling such two factors from partial point clouds enhances feature representation learning. To this end, our framework consists of three main parts: 1) a completion network to capture holistic semantics of point clouds; 2) a pose regression network to understand the viewing angle where partial data is scanned from; 3) a partial reconstruction network to encourage the model to learn content and pose features. To demonstrate the robustness of the learnt feature representations, we conduct several downstream tasks including classification, part segmentation, and registration, with comparisons against state-of-the-art methods. Our method not only outperforms existing self-supervised methods, but also shows a better generalizability across synthetic and real-world datasets.