Created by Li (Eric) Yi, Wang Zhao, He Wang, Minhyuk Sung, Leonidas J. Guibas.

If you find our work useful in your research, please consider citing:

    @article{yi2018gspn,
      title={Gspn: Generative shape proposal network for 3d instance segmentation in point cloud},
      author={Yi, Li and Zhao, Wang and Wang, He and Sung, Minhyuk and Guibas, Leonidas},
      journal={arXiv preprint arXiv:1812.03320},
      year={2018}
    }

This work is based on our CVPR'19 paper. You can find arXiv version of the paper here. We introduce a 3D object proposal approach named Generative Shape Proposal Network (GSPN) for instance segmentation in point cloud data. We incorporate GSPN into a novel 3D instance segmentation framework named Region-based PointNet (R-PointNet) which allows flexible proposal refinement and instance segmentation generation.

In this repository we release code and pre-trained model for both GSPN and R-PointNet.

We provide a step-by-step usage instruction from data processing to network evaluation on the ScanNet dataset. Please download ScanNet data and store in the "data" folder first.

1. Compiling the TF operators

    cd tf_ops
    . ./tf_all_compile.sh
   

2. Data pre-processing: convert ScanNet into downsampled point cloud for fast training and evaluation

    python data_prep.py ./data/scannet ./data/scannet_preprocessed
    python dataset.py
   

3. Two-step training: we train GSPN and R-PointNet separately in two different stages

    python train.py --train_module SPN --log_dir log_spn
    python train.py --train_module RPOINTNET --log_dir log_rpointnet --restore_model_path log_spn/model.ckpt --restore_scope shape_proposal_net
   

4. Evaluation on the validation set: we evaluation on downsampled point cloud from ScanNet validation scenes, where we first generate predictions and then evaluate based upon the official code provided by ScanNet Benchmark

    python test.py --model_path log_rpointnet/model.ckpt
    cd ./external/ScanNet/BenchmarkScripts/3d_evaluation
    python evaluate_semantic_instance.py --pred_path root_dir/eva/pred --gt_path root_dir/eva/gt --output_file root_dir/eva/pred/semantic_instance_evaluation.txt
   

The pretrained R-PointNet could be downloaded here (57Mb).

Our code is released under MIT License (see LICENSE file for details).