SGC: Efficient Semantic Scene Completion Network with Spatial Group Convolution

Created by Jiahui Zhang, Hao Zhao, Anbang Yao, Yurong Chen, Li Zhang and Hongen Liao

Citation

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

    @inproceedings{zhang2018efficient,
      title={Efficient Semantic Scene Completion Network with Spatial Group Convolution},
      author={Zhang, Jiahui and Zhao, Hao and Yao, Anbang and Chen, Yurong and Zhang, Li and Liao, Hongen},
      booktitle={European Conference on Computer Vision (ECCV)},
      year={2018}
    }

Introduction

This work is based on our ECCV'18 paper. You can find the paper here for a quick overview. SGC is designed for accelerating the computation of 3D dense prediction tasks. We conduct experiments on the SUNCG dataset, achieving state-of-the-art performance (84.5% of IoU for scene completion and 70.5% IoU for semantic scene completion) and fast speed.

In this repository we release code on SUNCG dataset.

Installation

0. Install SparseConvNet. This is a modified version of SparseConvNet. So you need to compile it yourself.

   Install the required packages:

   pip install torch==0.3.1
   pip install git+https://github.com/pytorch/tnt.git@master
   pip install msgpack
   pip install msgpack_numpy
   pip install cffi
   sudo apt-get install libsparsehash-dev
   pip install matplotlib
   

   Compile SparseConvNet:

   cd Pytorch
   python setup.py develop

1. Install SUNCG data toolbox. We provide a python wrapper for the C++ functions about SUNCG Dataset.

   Compile SUNCG data toolbox. We have tested on boost 1.58.0 and python2

   cd ssc/suncg_data_tools
   mkdir build
   cd build
   cmake ..
   make

Generate data for training and testing

0. Download SUNCG data, refer to SSCNet.

   cd ssc/
   mkdir data
   wget http://sscnet.cs.princeton.edu/sscnet_release/data/depthbin_eval.zip
   unzip depthbin_eval.zip
   wget http://sscnet.cs.princeton.edu/sscnet_release/data/SUNCGtrain.zip
   unzip SUNCGtrain.zip
   mv SUNCGtrain* depthbin/

1. Prepare data used in our project. It will take a long time and generate about 700G data.

   cd ssc/suncg_data_tools/script
   python prepare_data.py
   python prepare_weight.py

Usage

Pretrained model are provided in ssc/baseline/log and ssc/sgc-pattern4/log.

0. For baseline network without using SGC:

   cd ssc/baseline
   python sscnet

1. For network with SGC:

   cd ssc/sgc-pattern4
   python sscnet