Point-BERT: Pre-training 3D Point Cloud Transformers with Masked Point Modeling

TL;DR

Point-BERT introduces a BERT-style pre-training framework for 3D point cloud Transformers by learning a discrete token vocabulary via a dVAE and applying a Masked Point Modeling objective. It leverages local patch tokens, a standard Transformer backbone, and a Patch Mixing + MoCo setup to capture both local geometry and high-level semantics. The approach achieves state-of-the-art like performance on ModelNet40 and ScanObjectNN and demonstrates strong cross-domain and few-shot transfer, using a pure Transformer with minimal inductive bias. This work demonstrates the viability of unified Transformer-based representations for both 2D and 3D vision through self-supervised pre-training.

Abstract

We present Point-BERT, a new paradigm for learning Transformers to generalize the concept of BERT to 3D point cloud. Inspired by BERT, we devise a Masked Point Modeling (MPM) task to pre-train point cloud Transformers. Specifically, we first divide a point cloud into several local point patches, and a point cloud Tokenizer with a discrete Variational AutoEncoder (dVAE) is designed to generate discrete point tokens containing meaningful local information. Then, we randomly mask out some patches of input point clouds and feed them into the backbone Transformers. The pre-training objective is to recover the original point tokens at the masked locations under the supervision of point tokens obtained by the Tokenizer. Extensive experiments demonstrate that the proposed BERT-style pre-training strategy significantly improves the performance of standard point cloud Transformers. Equipped with our pre-training strategy, we show that a pure Transformer architecture attains 93.8% accuracy on ModelNet40 and 83.1% accuracy on the hardest setting of ScanObjectNN, surpassing carefully designed point cloud models with much fewer hand-made designs. We also demonstrate that the representations learned by Point-BERT transfer well to new tasks and domains, where our models largely advance the state-of-the-art of few-shot point cloud classification task. The code and pre-trained models are available at https://github.com/lulutang0608/Point-BERT

Figures (6)

• Figure 1: Illustration of our main idea. Point-BERT is designed for pre-training of standard point cloud Transformers. By training a dVAE via point cloud reconstruction, we can convert a point cloud into a sequence of discrete point tokens. Then we are able to pre-train the Transformers with a Mask Point Modeling (MPM) task by predicting the masked tokens.
• Figure 2: Masked point clouds reconstruction using our Point-BERT model trained on ShapeNet. We show the reconstruction results of synthetic objects from ShapeNet test set with block masking and random masking in the first two groups respectively. Our model also generalize well to unseen real scans from ScanObjectNN (the last two groups).
• Figure 3: The pipeline of Point-BERT. We first partition the input point cloud into several point patches (sub-clouds). A mini-PointNetpointnet is then used to obtain a sequence of point embeddings. Before pre-training, a Tokenizer is learned through dVAE-based point cloud reconstruction (as shown in the right part of the figure), where a point cloud can be converted into a sequence of discrete point tokens; During pre-training, we mask some portions of point embeddings and replace them with a mask token. The masked point embeddings are then fed into the Transformers. The model is trained to recover the original point tokens, under the supervision of point tokens obtained by the Tokenizer. We also add an auxiliary contrastive learning task to help the Transformers to capture high-level semantic knowledge.
• Figure 4: Visualization of feature distributions. We show the t-SNE visualization of feature vectors learned by Point-BERT (a) after pre-training, (b) after fine-tuning on ModelNet40, and (c) after fine-tuning on ScanObjectNN.
• Figure 5: Learning curve. We compare the performance of Transformers training from scratch (blue) and pre-training with Point-BERT (red) in terms of training loss and validation accuracy on synthetic and real-world object classification datasets.