Point-PEFT: Parameter-Efficient Fine-Tuning for 3D Pre-trained Models

Official implementation of 'Point-PEFT: Parameter-Efficient Fine-Tuning for 3D Pre-trained Models'.

The paper has been accepted by AAAI 2024.

[2023.5] We release ICCV2023 'ViewRefer3D', a multi-view framework for 3D visual grounding exploring how to grasp the view knowledge from both text and 3D modalities with LLM.

[2024.4] We release 'Any2Point', adapting Any-Modality pre-trained Models with 1% parameters to 3D downstream tasks with SOTA performance.

Introduction

We propose the Point-PEFT, a novel framework for adapting point cloud pre-trained models with minimal learnable parameters. Specifically, for a pre-trained 3D model, we freeze most of its parameters, and only tune the newly added PEFT modules on downstream tasks, which consist of a Point-prior Prompt and a Geometry-aware Adapter. The Point-prior Prompt constructs a memory bank with domain-specific knowledge and utilizes a parameter-free attention for prompt enhancement. The Geometry-aware Adapter aims to aggregate point cloud features within spatial neighborhoods to capture fine-grained geometric information.

Main Results

Comparison with existing 3D pre-trained models on the PB-T50-RS split of ScanObjectNN:

Method	Parameters	PB-T50-RS
Point-BERT	22.1M	83.1%
+Point-PEFT	0.6M	85.0%
Point-MAE-aug	22.1M	88.1%
+Point-PEFT	0.7M	89.1%
Point-M2AE-aug	12.9M	88.1%
+Point-PEFT	0.7M	88.2%

Comparison with existing 3D pre-trained models on the ModelNet40 without voting method:

Method	Parameters	Acc
Point-BERT	22.1M	92.7%
+Point-PEFT	0.6M	93.4%
Point-MAE	22.1M	93.2%
+Point-PEFT	0.8M	94.2%
Point-M2AE	15.3M	93.4%
+Point-PEFT	0.6M	94.1%

Ckpt Release

Real-world shape classification on the PB-T50-RS split of ScanObjectNN:

Method	Acc.	Logs
Point-M2AE-aug	88.2%	scan_m2ae.log
Point-MAE-aug	89.1%	scan_mae.log

Get Started

Installation

Create a conda environment and install basic dependencies:

git clone https://github.com/EvenJoker/Point-PEFT.git
cd Point-PEFT

conda create -n point-peft python=3.8
conda activate point-peft

# Install the according versions of torch and torchvision
conda install pytorch torchvision cudatoolkit
# e.g., conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3

pip install -r requirements.txt

Install GPU-related packages:

# Chamfer Distance and EMD
cd ./extensions/chamfer_dist
python setup.py install --user
cd ../emd
python setup.py install --user

# PointNet++
pip install "git+https://github.com/erikwijmans/Pointnet2_PyTorch.git#egg=pointnet2_ops&subdirectory=pointnet2_ops_lib"

# GPU kNN
pip install --upgrade https://github.com/unlimblue/KNN_CUDA/releases/download/0.2/KNN_CUDA-0.2-py3-none-any.whl

Dataset

For pre-training and fine-tuning, please follow DATASET.md to install ModelNet40, ScanObjectNN, and ShapeNetPart datasets, referring to Point-BERT. Specially Put the unzip folder under data/.

The final directory structure should be:

│Point-PEFT/
├──cfgs/
├──datasets/
├──data/
│   ├──ModelNet/
│   ├──ScanObjectNN/
├──...

Fine-tuning

M2AE:Please download the ckpt-best.pth, pre-train.pth and cache_shape.pt into the ckpts/ folder.

For the PB-T50-RS split of ScanObjectNN, run:

sh Finetune_cache_prompt_scan.sh

MAE:Please download the ckpt-best.pth, pre-train.pth and cache_shape.pt into the ckpts/ folder.

For the PB-T50-RS split of ScanObjectNN, run:

sh finetune.sh

Citation

If you find our paper and code useful in your research, please consider giving a star ⭐ and citation 📝.

@inproceedings{tang2024point,
  title={Point-PEFT: Parameter-efficient fine-tuning for 3D pre-trained models},
  author={Tang, Yiwen and Zhang, Ray and Guo, Zoey and Ma, Xianzheng and Zhao, Bin and Wang, Zhigang and Wang, Dong and Li, Xuelong},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={6},
  pages={5171--5179},
  year={2024}
}

Acknowledgement

This repo benefits from Point-M2AE, Point-BERT, Point-MAE. Thanks for their wonderful works.