Awesome

SGPN:Similarity Group Proposal Network for 3D Point Cloud Instance Segmentation [<a href="https://arxiv.org/abs/1711.08588">Arxiv</a>]

Dependencies

• tensorflow (1.3.0)
• h5py

Training & Testing

We firstly split the training set into training part and validation part. SGPN is finetuned on a pre-trained semantic segmentation model with large batchsize. For training,

python train.py 

Use the following scripts to generate results. valid.py is used to compute the per-category theshold for group merging. We then use <a href="github.com/ScanNet/ScanNet/blob/master/BenchmarkScripts/3d_evaluation/evaluate_semantic_instance.py">Scannet Evaluation</a> to evaluate test results.

python valid.py
python generate_results.py

Data and Model

Please refer to data/ for example h5 file and input list file. A pre-trained model can be downloaded [<a href="https://drive.google.com/file/d/1-e7YCfrLB4zqbFyWfQGe8sm_QFNrr59K/view?usp=sharing">here</a>].

Citation

If you find our work useful, please consider citing:

    @inproceedings{wang2018sgpn,
        title={SGPN: Similarity Group Proposal Network for 3D Point Cloud Instance Segmentation},
        author={Wang, Weiyue and Yu, Ronald and Huang, Qiangui and Neumann, Ulrich},
        booktitle={CVPR},
        year={2018}
    }

Acknowledgemets

This project is built upon [<a href="https://github.com/charlesq34/pointnet">PointNet</a>] and [<a href="https://github.com/charlesq34/pointnet2">PointNet++</a>].