Created by Saining Xie*, Sainan Liu*, Zeyu Chen, Zhuowen Tu from University of California, San Diego.

This repository provides a sample code for the paper Attentional ShapeContextNet for Point Cloud Recognition. In the paper, we introduce a neural network based algorithm by adopting the concept of shape context kernel for 3D shape recognition. The resulting network is ShapeContextNet (SCN), which has hierarchical modules that can represent the intrinsic property of object points by capturing and propagating both the local part and the global shape information. Additionally, we propose Attentional ShapeContextNet (A-SCN) which automate the process for the contextual region selection, feature aggregation, and feature transformation.

In this repository, we provide a sample code for A-SCN.

We use the same set of datasets from PointNet, and we have run our code in the following environment:

• python 3.6
• tensorflow 1.11.0
• CUDA 9.0
• cuDNN 7
• Ubuntu 16.04

To install h5py for Python:

sudo apt-get install libhdf5-dev
sudo pip install h5py

To run this code, we use a docker image that is built on top of nvidia/cuda:9.0-cudnn7-devel-ubuntu16.04, similar docker files can be found from this third-party repository.

For shape classification, part segmentation and semantic segmentation, please follow the instructions under the classification, part_seg and sem_seg folders respectively.

Part of this code is built on top of PointNet / PointNet++ .

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

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

    @article{saining2018ascn,
      title={Attentional ShapeContextNet for Point Cloud Recognition},
      author={Xie, Saining and Liu, Sainan and Chen, Zeyu and Tu, Zhuowen},
      year={2018}
    }