GVCNN (Group-View Convolutional Neural Networks for 3D Shape Recognition)

Data

• Download modelnet10-Class Orientation-aligned Subset
  • make .png images in order below
    • data_utils/make_views_dir.py
    • data_utils/off2obj.py (after 'sudo apt install openctm-tools')
    • data_utils/obj2png.py
  • Or You can create 2D dataset from 3D objects (.obj, .stl, and .off), using BlenderPhong.
• Or Downsized modelnet40(from https://drive.google.com/file/d/0B4v2jR3WsindMUE3N2xiLVpyLW8/view) to modelnet12/6-view.

Quick Start

• make group-view image tfrecord file
  • dataset_tools/create_modelnet_tf_record.py
• train.py

TODO

• balanced sampler

References from

• http://openaccess.thecvf.com/content_cvpr_2018/papers/Feng_GVCNN_Group-View_Convolutional_CVPR_2018_paper.pdf
• https://github.com/WeiTang114/MVCNN-TensorFlow
• https://github.com/pclausen/obj2png