By Liyao Tang, Zhe Chen, and Dacheng Tao
This is the implementation of our NeurIPS 2025 paper:
On Geometry-Enhanced Parameter-Efficient Fine-Tuning for 3D Scene Segmentation
[arXiv]
[OpenReview]
[Poster & Video]
If you find our work useful in your research, please consider citing:
@inproceedings{tang2025on,
title={On Geometry-Enhanced Parameter-Efficient Fine-Tuning for 3D Scene Segmentation},
author={Liyao Tang and Zhe Chen and Dacheng Tao},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
year={2025},
url={https://openreview.net/forum?id=7Ck1qqojSL}
}
This code is derived from Pointcept and has been tested with Python 3.10, PyTorch 2.3, CUDA 11.8 on Ubuntu 16.04.
We have also updated the code and environment.yml to be compatible with Pointcept v1.6.
Follow the installation of Pointcept, or directly clone this repo to setup locally:
conda env create -f environment.yml --verbose
conda activate pointcept-torch2.5.0-cu12.4
We follow the data preparation of Poincept. The configs are provided under configs/peft to run with Pointcept training scripts.
Alternatively, one can simply run the following script to start the training:
python main.py -c path/to/config.py --mode train --gpus [num gpus]
For more options, please check the main.py as well as the config files to play around.
We provide training logs here for reference.
Note: please modify the path to data and the pretrained model weights in the config for successful running.
Simply provide the training command with mode=test and the path for finding saved model, i.e.
python main.py -c path/to/config.py --mode test --weight [path]
This project is primarily licensed under the MIT License, see LICENSE.
It builds upon and includes components adapted from the following projects:
We thank the authors of these excellent works.
Code derived from these projects remains subject to their original licenses, as specified in the respective upstream repositories.