PointCLIP: Point Cloud Understanding by CLIP

Official implementation of 'PointCLIP: Point Cloud Understanding by CLIP'.

The paper has been accepted by CVPR 2022.

PointCLIP V2 with much stronger zero-shot performance will be released at repo.

Introduction

PointCLIP is the first to apply CLIP for point cloud recognition, which transfers 2D pre-trained knowledge into 3D domains. To achieve zero-shot classification, we encode a point cloud by projecting it onto multi-view depth maps and aggregate the view-wise prediction in an end-to-end manner. On top of that, we design an inter-view adapter to further enhance the few-shot performance, and explore the complementary property of PointCLIP for muti-knowledge ensemble.

Requirements

Installation

Create a conda environment and install dependencies:

git clone https://github.com/ZrrSkywalker/PointCLIP.git
cd PointCLIP

conda create -n pointclip python=3.7
conda activate pointclip

pip install -r requirements.txt

# Install the according versions of torch and torchvision
conda install pytorch torchvision cudatoolkit

# Install the modified dassl library (no need to re-build if the source code is changed)
cd Dassl3D/
python setup.py develop

cd ..

Dataset

Download the official ModelNet40 dataset and put the unzip folder under data/. The directory structure should be:

│PointCLIP/
├──...
├──data/
│   ├──modelnet40_ply_hdf5_2048/
├──...

Get Started

Zero-shot PointCLIP

Edit the running settings in scripts/zeroshot.sh, e.g. config file and output directory. Then run Zero-shot PointCLIP:

cd scripts
bash zeroshot.sh

If you need the post-search for the best view weights, add --post-search and modulate the search parameters in the config file. More search time leads to higher search results but longer time.

Few-shot PointCLIP

Set the shot number and other settings in scripts/fewshot.sh. Then run PointCLIP with the inter-view adapter:

cd scripts
bash fewshot.sh

--post-search is also optional.

Evaluation

Download the pre-pretrained checkpoint by 16-shot fine-tuning and put it under ckpt/adapter/. It will produce 86.71% on ModelNet40 test set and 87%+ by post-search:

cd scripts
bash eval.sh

You can edit the --model-dir and --output-dir to evaluate checkpoints trained by your own.

Acknowlegment

This repo benefits from CLIP, SimpleView and the excellent codebase Dassl. Thanks for their wonderful works.

Citation

@article{zhang2021pointclip,
  title={PointCLIP: Point Cloud Understanding by CLIP},
  author={Zhang, Renrui and Guo, Ziyu and Zhang, Wei and Li, Kunchang and Miao, Xupeng and Cui, Bin and Qiao, Yu and Gao, Peng and Li, Hongsheng},
  journal={arXiv preprint arXiv:2112.02413},
  year={2021}
}

Contact

If you have any question about this project, please feel free to contact [email protected].