• training code
• evaluation code
• evaluation data
• training data
• code for data engine

We will provide the training datasets as soon as posible.

Our model support mesh segmentation by sampling points from the mesh and propagate segmentation labels back to the mesh. We deployed a huggingface demo at 🤗demo🤗. We also provide local mesh segmentation demo at https://github.com/zyc00/point-sam-demo. Some meshes in glb format are provided in examples.

The code requires python>=3.8, timm>=0.9.0, pytorch>=2.1.0, torchvision>=0.16.0. Please follow the offcial guide to install Pytorch, timm and TorchVision dependencies. We also recommend compiling third partiy modules with g++=9.3.0.

Install third party modules.

# Install g++=9.3.0 in conda environment by
conda install gxx_linux-64=9.3.0
# Install torkit3d
git submodule update --init third_party/torkit3d && 
FORCE_CUDA=1 pip install third_party/torkit3d &&
# Install apex
git submodule update --init third_party/apex &&
pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" third_party/apex

We use accelerate as our training framework. The training scripts are in scripts. For example, if you want to train Point-SAM with ViT-large encoder, run bash scripts/train_large.sh.

We provide pretrained checkpoint for Point-SAM with ViT-L on HuggingFace. After downloading the checkpoint, you can inference with the following codes. For different numbers of points, we allow custom group_number and group_size. The default setting is group_number=512 and group_size=64, and we suggest setting group_number=2048 and group_size=256 while the number of points > 100k.

The evaluation code for KITTI360 is provided in evaluation/eval_kitti.py. You can run it by python eval_kitti.py --config large --ckpt_path your_ckpt_path. You can find the evaluation datasets in AGILE3D. We appreciate AGILE3D for providing evaluation datasets.

If you want to inference with your own point clouds and prompt points, see the codes in evaluation/inference.py.

We only support meshes in GLB format. If your mesh is not in GLB format, please convert it first. For our demo, you can simply drag a mesh into the browser, and the backend will encode the point cloud. Once encoding is complete, the camera view will adjust for optimal viewing. Then, you can interact with the demo by clicking on the mesh surface.

The performance of the Hugging Face demo is limited by data transfer and a shared GPU backend. To experience better performance, try running our local demo on your device by executing python app.py.

We provide a local demo for Point-SAM in demo. Some point cloud examples are provided in demo/static/models. If you want to try our demo, first install backend dependencies.

pip install flask flask-cors

Run the backend with following commands, you can change --host and --port to your IP address.

python demo/app.py --host localhost --port 5000 --pointcloud scene.ply --checkpoint ./pretrained/model.safetensors
# python demo/app.py --host localhost --port 5000 --pointcloud rhino.ply --checkpoint ./pretrained/model.safetensors
# python demo/app.py --host localhost --port 5000 --pointcloud tiko_10000_points.ply --checkpoint ./pretrained/model.safetensors
# python demo/app.py --host localhost --port 5000 --pointcloud tiko_50000_points.ply --checkpoint ./pretrained/model.safetensors

We refer to SAM, Uni3D and OpenShape while developing Point-SAM. Thanks for these awesome works. We also thank @YouJiacheng for providing constructive suggestions for the data engine design.

@misc{zhou2024pointsampromptable3dsegmentation,
      title={Point-SAM: Promptable 3D Segmentation Model for Point Clouds}, 
      author={Yuchen Zhou and Jiayuan Gu and Tung Yen Chiang and Fanbo Xiang and Hao Su},
      year={2024},
      eprint={2406.17741},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
      url={https://arxiv.org/abs/2406.17741}, 
}