RSNet Save

This is the official implementation of RSNet.

Project README


This is the official inplementation of Recurrent Slice Networks for 3D Segmentation on Point Clouds (RSNet), which is going to appear in CVPR 2018.

RSNet is a powerful and conceptually simple network for 3D point cloud segmentation tasks. It is fast and memory-efficient. In this repository, we release codes for training a RSNet on the S3DIS segmentation dataset. Training on other datasets can be easily achieved by following the same process.


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

        title={Recurrent Slice Networks for 3D Segmentation on Point Clouds},
        author={Huang, Qiangui and Wang, Weiyue and Neumann, Ulrich},
        journal={arXiv preprint arXiv:1802.04402},


  • python (tested on python2.7)
  • PyTorch (tested on 0.3.0)
  • cffi
  • h5py


  1. Clone this repository.
  2. Compile source codes for slice pooling/unpooling layers by following the readme file in layers

Data Preparation

  1. Process the S3DIS dataset by following the readme file in data.


  1. Launch training by the command below:
$ python

Type python --help for detailed input options. Be default, it will start the training by using Area 5 as testing set and others as training set.

During training, visualizations (.obj files) of intermediate predictions will be dumped into the folder results after each epoch. And they will be evaluated and saved in test_log.txt.


Codes in this repository are released under MIT License (see LICENSE file for details).

Open Source Agenda is not affiliated with "RSNet" Project. README Source: qianguih/RSNet
Open Issues
Last Commit
2 years ago

Open Source Agenda Badge

Open Source Agenda Rating