Upright orientation of 3D shapes with Convolutional Networks
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
data
util
voxelization
README.md
main.m
regression.prototxt
regression_fourlegged.caffemodel

README.md

Upright Orientation of 3D Shapes with Convolutional Networks

Introduction

  • This is the test code for

Zishun Liu, Juyong Zhang, Ligang Liu. Upright Orientation of 3D Shapes with Convolutional Networks. Graphical Models, 85: 22-29, 2016.

  • We have tested the code on Debian 8 and Matlab R2014b.
  • If you have any questions, please contact Zishun Liu via liuzishun@gmail.com.

Code

  • The root folder contains a trained model and interfaces for testing. The regression network for four-legged/wheeled group in the paper is provided.
  • The folder "data" contains several mesh files sampled from our test set, whose upright orientations are all positive z-axis.
  • The folder "util" is for utilities such as mesh loading and random rotation generation.
  • The folder "voxelization" is a toolbox to convert mesh models to their volume representations, from Jianxiong Xiao's Princeton Vision and Robotics Toolkit.

Usage

  1. Build Caffe (ND convolution is required) and its Matlab interface MatCaffe. Please refer to the official instructions 1 and 2.

  2. Compile the C-coded voxelization function in Matlab with mex ./voxelization/polygon2voxel_double.c.

  3. Edit the parameters in main.m and run it in Matlab. The results like the following would be printed:

     The prediction error is 2.7 degrees
    

Thank you!