PointSegBase is a code base for point cloud segmentation. It is a comprehensive and easy-to-use code base for point cloud segmentation, including various models, datasets, and evaluation metrics.

NEWS

• [2025/06/26] 🚀 Our work 'Controllable 3D Outdoor Scene Generation via Scene Graphs' is accepted by ICCV 2025.
• [2025/05/25] Basic code is released.
• [2025/02/05] 🚀 Our work 'Deep Hierarchical Learning for 3D Semantic Segmentation' is accepted by IJCV.
• [2024/08/12] 🚀 Our work 'Pyramid Diffusion for Fine 3D Large Scene Generation' is selected as Oral Presentation.
• [2024/07/02] 🚀 Our work 'Pyramid Diffusion for Fine 3D Large Scene Generation' is accepted by ECCV 2024.
• [2023/11/29] Official repo is created, code will be released soon, access our Project Page for more details.

Related Projects

There are several works that are based / used PointSegBase, including:

• 📌 Deep Hierarchical Learning for 3D Semantic Segmentation
  Chongshou Li, Yuheng Liu, Xinke Li, Yuning Zhang, Tianrui Li, Junsong Yuan
  International Journal of Computer Vision (IJCV) 2025
  [ Project ] [ Paper ]

• Controllable 3D Outdoor Scene Generation via Scene Graphs
  Yuheng Liu, Xinke Li, Yuning Zhang, Lu Qi, Xin Li, Wenping Wang, Chongshou Li, Xueting Li, Ming-Hsuan Yang
  International Conference on Computer Vision (ICCV) 2025
  [ Project ] [ arXiv ] [ Code ]

• Pyramid Diffusion for Fine 3D Large Scene Generation
  Yuheng Liu, Xinke Li, Xueting Li, Lu Qi, Chongshou Li, Ming-Hsuan Yang
  European Conference on Computer Vision (ECCV) 2024 - Oral Presentation (top 2%)
  [ Project ] [ arXiv ] [ Code ]

Citations

If you find this code base useful, please consider citing:

    @article{li2025deephierarchicallearningfor3dsemanticsegmentation,
        title={Deep Hierarchical Learning for 3D Semantic Segmentation},
        author={Chongshou Li and Yuheng Liu and Xinke Li and Yuning Zhang and Tianrui Li and Junsong Yuan},
        booktitle={International Journal of Computer Vision},
        year={2025}
    }

    @article{liu2025controllable3doutdoorscene,
        title={Controllable 3D Outdoor Scene Generation via Scene Graphs},
        author={Yuheng Liu and Xinke Li and Yuning Zhang and Lu Qi and Xin Li and Wenping Wang and Chongshou Li and Xueting Li and Ming-Hsuan Yang},
        booktitle={arXiv preprint arXiv:2503.07152},
        year={2025}
    }

    @InProceedings{liu2024pyramiddiffusionfine3d,
        title={Pyramid Diffusion for Fine 3D Large Scene Generation},
        author={Yuheng Liu and Xinke Li and Xueting Li and Lu Qi and Chongshou Li and Ming-Hsuan Yang},
        booktitle={European Conference on Computer Vision (ECCV)},
        year={2024},
    }

Installation

# Installation (Ubuntu 22.04 + CUDA 11.8)

# 1. Create a new conda environment
conda create -n PointSegBase python=3.10 -y
conda activate PointSegBase

# 2. Install PyTorch (CUDA 11.8, version < 2.0.0)
pip install torch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1

# 3. Install required packages
conda install -y pyyaml six numba h5py
pip install numpy==1.25.2 scipy==1.13.0 open3d

# 4. Install TorchSparse (choose one method)

# Option A: Install via official script
python -c "$(curl -fsSL https://raw.githubusercontent.com/mit-han-lab/torchsparse/master/install.py)"

# Option B: Install directly from GitHub
sudo apt-get install libsparsehash-dev
pip install git+https://github.com/mit-han-lab/torchsparse.git

Datasets

We provide the following datasets:

• Campus3D, Urban3D, PartNet-H
  [ Hugging Face ] [ Baidu Netdisk (Until 05/25/2026) ]

All datasets have been preprocessed and should be placed in the data folder.

Training

Before starting any training, you need to set the corresponding config file located in the configs folder. Use -c to specify the config path for the current experiment group, and --exp_name to set the path where logs and checkpoints for the current experiment group will be saved. We provide example configs for both single-layer and multi-layer training on the Campus3D and Urban3D datasets using UNet and PointNet++ networks, as well as for the Bed category in the PartNet-H dataset using the same networks. You can refer to these examples to create additional training configs for other categories in PartNet or for other network architectures.

export CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES}
python run.py -c ${CONFIG_PATH} --exp_name ${EXP_NAME}

Campus3D

# Settings (Campus3D, PointNet++, layer 1)
python run.py -c configs/pointnet_campus_layer_1 --exp_name exp_campus3d_pointnet_layer_1

# Settings (Campus3D, PointNet++, layer 2)
python run.py -c configs/pointnet_campus_layer_2 --exp_name exp_campus3d_pointnet_layer_2

# Settings (Campus3D, PointNet++, layer 3)
python run.py -c configs/pointnet_campus_layer_3 --exp_name exp_campus3d_pointnet_layer_3

# Settings (Campus3D, PointNet++, layer all)
python run.py -c configs/pointnet_campus_layer_all --exp_name exp_campus3d_pointnet_layer_all

# Settings (Campus3D, UNet, layer 1)
python run.py -c configs/unet_campus_layer_1 --exp_name exp_campus3d_unet_layer_1

# Settings (Campus3D, UNet, layer 2)
python run.py -c configs/unet_campus_layer_2 --exp_name exp_campus3d_unet_layer_2

# Settings (Campus3D, UNet, layer all)
python run.py -c configs/unet_campus_layer_all --exp_name exp_campus3d_unet_layer_all

Urban3D

# Settings (Urban3D, PointNet++, layer 1)
python run.py -c configs/pointnet_urban_layer_1 --exp_name exp_urban3d_pointnet_layer_1

# Settings (Urban3D, PointNet++, layer 2)
python run.py -c configs/pointnet_urban_layer_2 --exp_name exp_urban3d_pointnet_layer_2

# Settings (Urban3D, PointNet++, layer all)
python run.py -c configs/pointnet_urban_layer_all --exp_name exp_urban3d_pointnet_layer_all

# Settings (Urban3D, UNet, layer 1)
python run.py -c configs/unet_urban_layer_1 --exp_name exp_urban3d_unet_layer_1

# Settings (Urban3D, UNet, layer 2)
python run.py -c configs/unet_urban_layer_2 --exp_name exp_urban3d_unet_layer_2

# Settings (Urban3D, UNet, layer all)
python run.py -c configs/unet_urban_layer_all --exp_name exp_urban3d_unet_layer_all

PartNet-H

# Settings (PartNet-H-Bed, PointNet++, layer 1)
python run.py -c configs/pointnet_partnet-Bed_layer_1 --exp_name exp_partnet-Bed_pointnet_layer_1

# Settings (PartNet-H-Bed, PointNet++, layer 2)
python run.py -c configs/pointnet_partnet-Bed_layer_2 --exp_name exp_partnet-Bed_pointnet_layer_2

# Settings (PartNet-H-Bed, PointNet++, layer 3)
python run.py -c configs/pointnet_partnet-Bed_layer_3 --exp_name exp_partnet-Bed_pointnet_layer_3

# Settings (PartNet-H-Bed, PointNet++, layer all)
python run.py -c configs/pointnet_partnet-Bed_layer_all --exp_name exp_partnet-Bed_pointnet_layer_all

# Settings (PartNet-H-Bed, UNet, layer 1)
python run.py -c configs/unet_partnet-Bed_layer_1 --exp_name exp_partnet-Bed_unet_layer_1

# Settings (PartNet-H-Bed, UNet, layer 2)
python run.py -c configs/unet_partnet-Bed_layer_2 --exp_name exp_partnet-Bed_unet_layer_2

# Settings (PartNet-H-Bed, UNet, layer 3)
python run.py -c configs/unet_partnet-Bed_layer_3 --exp_name exp_partnet-Bed_unet_layer_3

# Settings (PartNet-H-Bed, UNet, layer all)
python run.py -c configs/unet_partnet-Bed_layer_all --exp_name exp_partnet-Bed_unet_layer_all

Evaluation

After completing the training of your model, you can use --eval to evaluate the trained model. Specifically, you need to go into the folder corresponding to the experiment name you previously set, and edit the train_setting.yaml file in the configs folder. Set IS_PRETRAINED to True and PRETRAINED_MODEL_PATH to the path of the model inside the checkpoints folder.

python run.py -c ${EXP_NAME}/configs --exp_name ${EVAL_NAME} --eval

Example

python run.py -c exp_campus3d_pointnet_layer_1/configs --exp_name exp_campus3d_pointnet_layer_1_eval --eval

Contact

This repository will be maintained by Yuheng. For any issues regarding code, dataset downloads, or the official website, you can contact Yuheng at the following email addresses: yuhengliu02@gmail.com.