PointLoRA

Song Wang, Xiaolu Liu, Lingdong Kong, Jianyun Xu, Chunyong Hu, Gongfan Fang, Wentong Li, Jianke Zhu, Xinchao Wang

This is the official implementation of PointLoRA: Low-Rank Adaptation with Token Selection for Point Cloud Learning (CVPR 2025) [Paper].

<p align="center"> <a><img src="fig/framework.png" width="90%"></a> </p>

Preparation

Environment Setup

We release the PointLoRA implementation with Point-MAE, please refer the environment setup in the original repo.

Dataset Download

We use ScanObjectNN, ModelNet40, and ShapeNetPart in this work. Please refer the data processing in Point-BERT.

Fine-tuning on Downstream Tasks

Fine-tuning the Point-MAE model with our proposed PointLoRA:

# For fine-tuning on PB-T50-RS variant
python main.py --config cfgs/finetune_scan_hardest_pointlora.yaml --ckpts <path/to/pre-trained/model> --finetune_model --exp_name pointlora_finetune

Acknowledgement

We gratefully acknowledge the contributions of various open-source projects that supported this work: Point-BERT, Point-MAE, DAPT, PPT, Point-PEFT.