SCPNet: Semantic Scene Completion on Point Cloud (CVPR 2023, Highlight)

News

• 2023-05 Preliminary codes are released.
• 2023-02 Our SCPNet is accepted by CVPR 2023 (Highlight)!
• 2022-11 Our method ranks 1st in SemanticKITTI Semantic Scene Completion Challenge, with mIoU=36.7.
• SCPNet is comprised of a novel completion sub-network without an encoder-decoder structure and a segmentation sub-network obtained by replacing the cylindrical partition of Cylinder3D with conventional cubic partition.

Installation

Requirements

• PyTorch >= 1.10
• pyyaml
• Cython
• tqdm
• numba
• Numpy-indexed
• torch-scatter
• spconv (tested with spconv==1.0 and cuda==11.3)

Data Preparation

SemanticKITTI

./
├── 
├── ...
└── path_to_data_shown_in_config/
    ├──sequences
        ├── 00/           
        │   ├── velodyne/	
        |   |	├── 000000.bin
        |   |	├── 000001.bin
        |   |	└── ...
        │   └── labels/ 
        |       ├── 000000.label
        |       ├── 000001.label
        |       └── ...
        │   └── voxels/ 
        |       ├── 000000.bin
        |       ├── 000000.label
        |       ├── 000000.invalid
        |       ├── 000000.occluded
        |       ├── 000001.bin
        |       ├── 000001.label
        |       ├── 000001.invalid
        |       ├── 000001.occluded
        |       └── ...
        ├── 08/ # for validation
        ├── 11/ # 11-21 for testing
        └── 21/
	    └── ...

Test

We take evaluation on the SemanticKITTI test set (single-scan) as example.

1. Download the pre-trained models and put them in ./model_load_dir.
2. Set val_data_loader>imageset: “test” in the configuration file config/semantickitti-multiscan.yaml.
3. Generate predictions on the SemanticKITTI test set.

CUDA_VISIBLE_DEVICES=0 python -u test_scpnet_comp.py

The model predictions will be saved in ./out_scpnet/test by default.

Train

1. Set val_data_loader>imageset: “test” in the configuration file config/semantickitti-multiscan.yaml.
2. train the network by running the train script

CUDA_VISIBLE_DEVICES=0 python -u train_scpnet_comp.py

Citation

If you use the codes, please cite the following publication:

@inproceedings{scpnet,
    title     = {SCPNet: Semantic Scene Completion on Point Cloud},
    author    = {Xia, Zhaoyang and Liu, Youquan and Li, Xin and Zhu, Xinge and Ma, Yuexin and Li, Yikang and Hou, Yuenan and Qiao, Yu},
    booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},
    year      = {2023}
}

Acknowledgements

We thanks for these codebases, including Cylinder3D, PVKD and spconv.