Skip to content

suhangpro/mvcnn

Repository files navigation

Multi-view CNN (MVCNN) for shape recognition

Project Page MVCNN pipeline

The goal of the project is to learn a general purpose descriptor for shape recognition. To do this we train discriminative models for shape recognition using convolutional neural networks (CNNs) where view-based shape representations are the only cues. Examples include line-drawings, clip art images where color is removed, or renderings of 3D models where there is little or no texture information present.

If you use any part of the code from this project, please cite:

@inproceedings{su15mvcnn, author = {Hang Su and Subhransu Maji and Evangelos Kalogerakis and Erik G. Learned{-}Miller}, title = {Multi-view convolutional neural networks for 3d shape recognition}, booktitle = {Proc. ICCV}, year = {2015}}

Other implementations

(These are implementations provided by friends or found online, and are listed here for your convenience. I do not provide direct support on them.)

Installation

  • Install dependencies
git submodule update --init
  • Compile

compile for CPU:

# two environment variables might need to be set, e.g. MATLABDIR=<MATLAB_ROOT> MEX=<MATLAB_ROOT>/bin/mex
matlab -nodisplay -r "setup(true);exit;"

compile for GPU (w/ cuDNN):

# 1) two environment variables might need to be set, e.g. MATLABDIR=<MATLAB_ROOT> MEX=<MATLAB_ROOT>/bin/mex
# 2) other compilation options (e.g. 'cudaRoot',<CUDA_ROOT>,'cudaMethod','nvcc','cudnnRoot',<CUDNN_ROOT>) 
#  might be needed in the 'struct(...)' as well depending on you system settings
matlab -nodisplay -r "setup(true,struct('enableGpu',true,'enableCudnn',true));exit;"

Note: you can alternatively run directly the scripts from the Matlab command window, e.g. for Windows installations: setup(true,struct('enableGpu',true,'cudaRoot','C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.0','cudaMethod','nvcc')); You may also need to add Visual Studio's cl.exe in your PATH environment (e.g., C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\bin\amd64)

Usage

  • Extract descriptor for a shape (.off/.obj mesh). The descriptor will be saved in a .txt file (e.g. bunny_descriptor.txt). Uses default model with no fine-tuning. Assumes upright orientation by default.
MATLAB> shape_compute_descriptor('bunny.off');
  • Extract descriptor for all shapes in a folder (.off/.obj meshes). The descriptors will be saved in .txt files in the same folder. Assumes no upright orientation.
MATLAB> shape_compute_descriptor('my_mesh_folder/','useUprightAssumption',false);
  • Extract descriptors for all shapes in a folder (.off/.obj meshes) and post-process descriptors with learned metric. Uses non-default models.
MATLAB> shape_compute_descriptor('my_mesh_folder/', 'cnnModel', 'my_cnn.mat', ...
'metricModel', 'my_metric.mat','applyMetric',true); 
  • Download datasets for training/evaluation (should be placed under data/)

    • modelnet40v1 (12 views w/ upright assumption): tarball (204M)
    • modelnet40v2 (80 views w/o upright assumption): tarball (1.3G)
    • shapenet55v1 (12 views w/ upright assumption): tarball (2.4G)
    • shapenet55v2 (80 views w/o upright assumption): tarball (15G)
  • Run training examples (see run_experiments.m for details)

# LD_LIBRARY_PATH might need to be set, e.g. LD_LIBRARY_PATH=<CUDA_ROOT>/lib64:<CUDNN_ROOT> 
matlab -nodisplay -r "run_experiments;exit;"