laughtervv/SGPN

SGPN:Similarity Group Proposal Network for 3D Point Cloud Instance Segmentation [Arxiv]

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 Scannet Evaluation 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 [here].

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 [PointNet] and [PointNet++].