PointMamba

A Simple State Space Model for Point Cloud Analysis

Dingkang Liang¹ *, Xin Zhou¹ *, Wei Xu¹, Xingkui Zhu¹, Zhikang Zou², Xiaoqing Ye², Xiao Tan² and Xiang Bai^1†

¹ Huazhong University of Science & Technology, ² Baidu Inc.

(*) Equal contribution. ($\dagger$) Corresponding author.

📣 News

[11/Oct/2024] 🚀 Check out our latest work PointGST which achieves 99.48%, 97.76%, and 96.18% overall accuracy on the ScanObjNN OBJ_BG, OBJ_OBLY, and PB_T50_RS datasets, respectively.
[26/Sept/2024] PointMamba is accepted to NeurIPS 2024! 🥳🥳🥳
[30/May/2024] Update! We update the architecture and performance. Please check our latest paper and compare it with the new results. Code and weight will be updated soon.
[01/Apr/2024] ScanObjectNN with further data augmentation is now available, check it out!
[16/Mar/2024] The configurations and checkpoints for ModelNet40 are now accessible, check it out!
[05/Mar/2024] Our paper DAPT (github) has been accepted by CVPR 2024! 🥳🥳🥳 Check it out and give it a star 🌟!
[16/Feb/2024] Release the paper.

Abstract

Transformers have become one of the foundational architectures in point cloud analysis tasks due to their excellent global modeling ability. However, the attention mechanism has quadratic complexity, making the design of a linear complexity method with global modeling appealing. In this paper, we propose PointMamba, transferring the success of Mamba, a recent representative state space model (SSM), from NLP to point cloud analysis tasks. Unlike traditional Transformers, PointMamba employs a linear complexity algorithm, presenting global modeling capacity while significantly reducing computational costs. Specifically, our method leverages space-filling curves for effective point tokenization and adopts an extremely simple, non-hierarchical Mamba encoder as the backbone. Comprehensive evaluations demonstrate that PointMamba achieves superior performance across multiple datasets while significantly reducing GPU memory usage and FLOPs. This work underscores the potential of SSMs in 3D vision-related tasks and presents a simple yet effective Mamba-based baseline for future research.

Overview

Main Results

The table below will be updated once the code for the latest version release is ready.

Task	Dataset	Config	Acc.(Scratch)	Download (Scratch)	Acc.(pretrain)	Download (Finetune)
Pre-training	ShapeNet	pretrain.yaml			N.A.	here
Classification	ModelNet40	finetune_modelnet.yaml	92.4%	here	93.6%	here
Classification	ScanObjectNN	finetune_scan_objbg.yaml	88.30%	here	90.71%	here
Classification*	ScanObjectNN	finetune_scan_objbg.yaml	\	\	93.29%	here
Classification	ScanObjectNN	finetune_scan_objonly.yaml	87.78%	here	88.47%	here
Classification*	ScanObjectNN	finetune_scan_objonly.yaml	\	\	91.91%	here
Classification	ScanObjectNN	finetune_scan_hardest.yaml	82.48%	here	84.87%	here
Classification*	ScanObjectNN	finetune_scan_hardest.yaml	\	\	88.17%	here
Part Segmentation	ShapeNetPart	part segmentation	85.8% mIoU	here	86.0% mIoU	here

* indicates further using simple rotational augmentation for training.

Getting Started

Datasets

See DATASET.md for details.

Usage

See USAGE.md for details.

To Do

Release code.
Release checkpoints.
ModelNet40.
Update the code.

Acknowledgement

This project is based on Point-BERT (paper, code), Point-MAE (paper, code), Mamba (paper, code), Causal-Conv1d (code). Thanks for their wonderful works.

Citation

If you find this repository useful in your research, please consider giving a star ⭐ and a citation

@inproceedings{liang2024pointmamba,
      title={PointMamba: A Simple State Space Model for Point Cloud Analysis}, 
      author={Liang, Dingkang and Zhou, Xin and Xu, Wei and Zhu, Xingkui and Zou, Zhikang and Ye, Xiaoqing and Tan, Xiao and Bai, Xiang},
      booktitle={Advances in Neural Information Processing Systems},
      year={2024}
}

LMD0311/PointMamba