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

By Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, Jiwen Lu. Computer Vision and Pattern Recognition, 2021

1,136 citations recorded by the source corpus.

We present Point-BERT, a new paradigm for learning Transformers to generalize the concept of BERT [8] 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.

Subfields: Computational Mechanics, Computer Graphics and Computer-Aided Design, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/dd2819016c6bf244c39b3e6707b60389bbdbcd21

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computational Mechanics

In Computer Graphics and Computer-Aided Design

In Computer Vision and Pattern Recognition

Citations in the corpus

A Survey of Visual Transformers (2021, methodology): “Speciﬁcally, Point-BERT [113] and Point-MAE [114] directly transfer the previous works [70], [71] to point clouds, while MaskPoint [115] changes the generative training scheme by using a contrastive decoder as similar as DINO (2022) [92] for binary noise/part classiﬁcation.”
Foundation models in robotics: Applications, challenges, and the future (2023, positive): “PointBERT [55] uses a transformer-based architecture to extract features from point clouds, generalizing the concept of BERT into 3D point clouds.”
A Survey on Masked Autoencoder for Self-supervised Learning in Vision and Beyond (2022, methodology): “Motivated by the success of BEiT in vision, [26] extends masked modeling strategy to point cloud with masked point modeling termed Point-BERT.”
Transformer-Based Visual Segmentation: A Survey (2023, positive): “PointBert [257] designs the ﬁrst Masked Point Modeling (MPM) task to pre-train point cloud transformers.”
MM-LLMs: Recent Advances in MultiModal Large Language Models (2024, neutral): “3 of Appendix G.”
Point Transformer V3: Simpler, Faster, Stronger (2023, background): “While major efforts in 3D representation learning focused on individual objects [57, 68, 69, 87, 103], some recent advancements have redirected attention towards training on real-world scene-centric point clouds [30, 36, 91, 94, 107].”
MeshGPT: Generating Triangle Meshes with Decoder-Only Transformers (2023, methodology): “Following recent works for unconditional shape generation [13, 66, 67] for calculating the shape metrics we define where in the 1-NNA metric N X is a point cloud that is closest to X in both generated and reference dataset, i.e., N X = argmin K ∈ S r ∪ S g D ( X, K ) We use a Chamfer Distance (CD)…”
Transolver: A Fast Transformer Solver for PDEs on General Geometries (2024, methodology): “Besides, 3D-GeoCA enhances GNO by employing the advanced 3D geometric deep model Point-BERT (Yu et al., 2022) as the feature encoder.”
PointMamba: A Simple State Space Model for Point Cloud Analysis (2024, methodology): “Specifically, Point-BERT [59] and Point-MAE [38] introduce a standard Transformer architecture for self-supervised learning and is applicable to different downstream tasks.”
PointLLM: Empowering Large Language Models to Understand Point Clouds (2023, methodology): “Point-BERT [52], pre-trained with ULIP-2 [50] on the Objaverse [9] dataset, serves as our point encoder.”
PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies (2022, positive): “Following Point-BERT [52], the number of input points is set to 1 , 024 , where the points are randomly sampled during training and uniformly sampled during testing (denoted as point resampled augmentation).”
Point-Bind & Point-LLM: Aligning Point Cloud with Multi-modality for 3D Understanding, Generation, and Instruction Following (2023, background): “In the last two columns, we utilize different 3D encoders in Point-Bind, i.e., Point-BERT [86], PointNeXt [57], and I2P-MAE [92].”
Masked Autoencoders for Point Cloud Self-supervised Learning (2022, methodology): “A simple method for Position Embedding (PE) is mapping coordinates of centers to embedding dimension with a learnable MLP, following previous works [54,59].”
ULIP: Learning a Unified Representation of Language, Images, and Point Clouds for 3D Understanding (2022, methodology): “Overall Acc Class-mean Acc PointNet [35] 68.2 63.4 PointNet++ [36] 77.9 75.4 DGCNN [49] 78.1 73.6 MVTN [15] 82.8 – PointBERT [58] 83.1 – RepSurf-U [40] 84.6 – PointMAE [33] 85.2 – RepSurf-U (2x) [40] 86.0 – [32], that shows superior performance as our image-text encoders.”