PointNeXt

[arXiv] | [OpenPoints Library] | [Online Documentation]

Official PyTorch implementation for the following paper:

PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies

by Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, Bernard Ghanem

TL;DR: We propose improved training and model scaling strategies to boost PointNet++ to the state-of-the-art level. PointNet++ with the proposed model scaling is named as PointNeXt, the next version of PointNets.

News

• 💥 Sep, 2022: PointNeXt accepted by NeurIPS'22
• 💥 Jun, 2022: Code released

Features

In the PointNeXt project, we propose a new and flexible codebase for point-based methods, namely OpenPoints. The biggest difference between OpenPoints and other libraries is that we focus more on reproducibility and fair benchmarking.

1. Extensibility: supports many representative networks for point cloud understanding, such as PointNet, DGCNN, DeepGCN, PointNet++, ASSANet, PointMLP, and our PointNeXt. More networks can be built easily based on our framework since OpenPoints support a wide range of basic operations including graph convolutions, self-attention, farthest point sampling, ball query, e.t.c.

2. Reproducibility: all implemented models are trained on various tasks at least three times. Mean±std is provided in the PointNeXt paper. Pretrained models and logs are available.

3. Fair Benchmarking: in PointNeXt, we find a large part of performance gain is due to the training strategies. In OpenPoints, all models are trained with the improved training strategies and all achieve much higher accuracy than the original reported value.

4. Ease of Use: Build model, optimizer, scheduler, loss function, and data loader easily from cfg. Train and validate different models on various tasks by simply changing the cfg\*\*.yaml file.

   model = build_model_from_cfg(cfg.model)
   criterion = build_criterion_from_cfg(cfg.criterion_args)
   

   Here is an example of pointnet.yaml (model configuration for PointNet model):

   model:
     NAME: BaseCls
     encoder_args:
       NAME: PointNetEncoder
       in_channels: 4
     cls_args:
       NAME: ClsHead
       num_classes: 15
       in_channels: 1024
       mlps: [512,256]
       norm_args: 
         norm: 'bn1d'
   

5. Online logging: Support wandb for checking your results anytime anywhere. Just set wandb.use_wandb=True in your command.

   

Installation

Pip packages

The Python libraries are released as installable packages:

pip install pointnext_official

pointnext_official installs openpoints as its core library dependency and provides PointNeXt release metadata and checkpoint download helpers, for example:

pointnext-download --list
pointnext-download modelnet40-pointnext-s-c64 --output-dir ./hf_cache

The PyPI packages are importable without compiling CUDA extensions. Full training/evaluation still requires the custom CUDA/C++ ops, so use a source checkout for benchmark reproduction.

Source install for training/evaluation

git clone --recurse-submodules https://github.com/guochengqian/PointNeXt.git
cd PointNeXt
git submodule update --init --recursive
source update.sh
source install.sh

If SSH is configured, git@github.com:guochengqian/PointNeXt.git also works. The current install.sh targets modern NVIDIA GPUs, including Blackwell (RTX 50-series / RTX 5090, sm_120), via a uv + Python 3.12 environment with PyTorch ≥2.7 built for CUDA 12.8 (cu128). Blackwell requires CUDA ≥12.8 and PyTorch ≥2.7; the legacy CUDA 11.3 recipe (kept commented at the bottom of install.sh) does not run on an RTX 5090. To compile the CUDA ops you also need a matching CUDA toolkit (nvcc) and a host compiler the toolkit supports (CUDA 12.x needs g++ < 14, e.g. sudo apt install gcc-13 g++-13). Modify install.sh for a different CUDA/PyTorch version. See Install, FAQ, and Checkpoints for details.

Usage

Check our online documentation for detailed instructions.

A short instruction: all experiments follow the simple rule to train and test:

CUDA_VISIBLE_DEVICES=$GPUs python examples/$task_folder/main.py --cfg $cfg $kwargs

• $GPUs is the list of GPUs to use, for most experiments (ScanObjectNN, ModelNet40, S3DIS), we only use 1 A100 (GPUs=0)
• $task_folder is the folder name of the experiment. For example, for s3dis segmentation, $task_folder=s3dis
• $cfg is the path to cfg, for example, s3dis segmentation, $cfg=cfgs/s3dis/pointnext-s.yaml
• $kwargs are the other keyword arguments to use. For example, testing in S3DIS area 5, $kwargs should be mode=test, --pretrained_path $pretrained_path.

Model Zoo (pretrained weights)

See Model Zoo and checkpoint download docs. The recommended new release path hosts large checkpoints and checksum manifests on Hugging Face Hub, while GitHub Releases/PyPI host source and Python packages.

Visualization

More examples are available in the paper.

Acknowledgment

This library is inspired by PyTorch-image-models and mmcv.

Citation

If you find PointNeXt or the OpenPoints codebase useful, please cite:

@InProceedings{qian2022pointnext,
  title   = {PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies},
  author  = {Qian, Guocheng and Li, Yuchen and Peng, Houwen and Mai, Jinjie and Hammoud, Hasan and Elhoseiny, Mohamed and Ghanem, Bernard},
  booktitle=Advances in Neural Information Processing Systems (NeurIPS),
  year    = {2022},
}