Self-supervised Learning of Rotation-invariant 3D Point Set Features using Transformer and its Self-distillation

TL;DR

This paper introduces a self-supervised learning framework using Transformer architecture to learn accurate, rotation-invariant features from unlabeled 3D point sets, addressing the challenge of label scarcity in 3D shape analysis.

Contribution

It proposes a novel lightweight DNN with self-attention and self-distillation for rotation-invariant 3D feature learning from unlabeled data, improving over existing supervised methods.

Findings

The proposed method learns more accurate rotation-invariant features.

Existing rotation-invariant architectures are less effective under self-supervision.

Combining multi-crop and cut-mix enhances training diversity.

Abstract

Invariance against rotations of 3D objects is an important property in analyzing 3D point set data. Conventional 3D point set DNNs having rotation invariance typically obtain accurate 3D shape features via supervised learning by using labeled 3D point sets as training samples. However, due to the rapid increase in 3D point set data and the high cost of labeling, a framework to learn rotation-invariant 3D shape features from numerous unlabeled 3D point sets is required. This paper proposes a novel self-supervised learning framework for acquiring accurate and rotation-invariant 3D point set features at object-level. Our proposed lightweight DNN architecture decomposes an input 3D point set into multiple global-scale regions, called tokens, that preserve the spatial layout of partial shapes composing the 3D object. We employ a self-attention mechanism to refine the tokens and aggregate them into an expressive rotation-invariant feature per 3D point set. Our DNN is effectively trained by using pseudo-labels generated by a self-distillation framework. To facilitate the learning of accurate features, we propose to combine multi-crop and cut-mix data augmentation techniques to diversify 3D point sets for training. Through a comprehensive evaluation, we empirically demonstrate that, (1) existing rotation-invariant DNN architectures designed for supervised learning do not necessarily learn accurate 3D shape features under a self-supervised learning scenario, and (2) our proposed algorithm learns rotation-invariant 3D point set features that are more accurate than those learned by existing algorithms. Code is available at https://github.com/takahikof/RIPT_SDMM