Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time


This code implements the method of deep learning local descriptors for 3D surface shapes described in the following paper:

  Hanyu Wang, Jianwei Guo, Dong-Ming Yan, Weize Quan, Xiaopeng Zhang. 
  Learning 3D Keypoint Descriptors for Non-Rigid Shape Matching. 
  ECCV 2018.

An advanced method is implemented in LPS.

Please consider citing the above paper if you use the code/program (or part of it).


This program is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version.


There are three folders here. The "cpp" calls the "matlab" for GI generation. The "python" is used for network training and testing.

The usage is as follows:

  1. Compile matlab project:

     MCC matlab "mcc -W cpplib:libcompcur -T link:lib compute_curvature.m". We got the libcompcur.dll that will be added to the CPP project.
  2. Build cpp solution: this code is to generate geometry images. You can run this step in your local desktop.

     Modify CMakeLists :
     	add include_directories and link_directories for openmesh and matlab runtime
     Build solution
     Modify config.ini for mesh_dir(directory of OFF models) gi_dir(directory of geometry images) and kpi_dir(directory of key points, you can skip it for dense matching)
     	edit other paras such as gi_size(NxN of gi), hks_len, rotation_num and radius_list_p (the ratio of geodesic diameter).
     Add "libcompcur.dll" to folder with GIGen.exe
     Run "GIGen.exe config.ini" to generate GI
  3. Python project: this code is to train and test network. You should copy the geometry images generated in last step into the server.

    -Train network:

     run "" to classify dataset by point index for generating Tfrecord
     	source_dir is the folder of geometry images (rotation_num=12) for training, destination_dir is the generated folder after splitting
     run "" to generate Tfrecord
     	gi_dir is the same as destination_dir, tfr_dir is the generated folder of Tfrecord.
     run "" to pretrain a classification network
     	tfr_dir needs to be specified
     run "" to train the triplet network by restoring a pre-trained classification model
     	Tfr_dir needs to be specified

    -Test to generate descriptor:

     run "" to generate descriptor using test dataset
     	gi_dir is the folder of geometry images (rotation_num=1) for testing, desc_dir is the generated folder of descriptor.


Should you have any questions, comments, or suggestions, please contact me at:

Jianwei Guo:

October, 2018

Copyright (C) 2018


Deep learning local descriptors for 3D surface shapes






No releases published


No packages published