Skip to content

angeladai/cnncomplete

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

51 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

cnncomplete

This repo contains code to train a volumetric deep neural network to complete partially scanned 3D shapes. More information can be found in our paper.

Data

Train/test data is available for download on our project website.

Code

Installation:

Training tasks use Torch7, with torch packages cudnn, cunn, torch-hdf5, xlua.

Matlab visualization of the isosurface in testing uses the matio package.

The shape synthesis code was developed under VS2013, and uses flann (included in external).

Training:

  • th train_class.lua -model epn-unet-class -save logs-epn-unet-class -train_data data/h5_shapenet_dim32_sdf/train_shape_voxel_data_list.txt -test_data data/h5_shapenet_dim32_sdf/test_shape_voxel_data_list.txt -gpu_index 0
  • For more options, see help: th train_class.lua -h or th train.lua -h
  • Trained models: trained_models.zip (700mb)

Testing:

  • th test.lua --model_path [path to model] --test_file sampledata/scan.h5 --output_path [path to output] --classifier_path [path to classifier model, only specify if using epn-class or epn-unet-class models]
  • For more options, see help: th test.lua -h

Citation:

@inproceedings{dai2017complete,
  title={Shape Completion using 3D-Encoder-Predictor CNNs and Shape Synthesis},
  author={Dai, Angela and Qi, Charles Ruizhongtai and Nie{\ss}ner, Matthias},
  booktitle = {Proc. Computer Vision and Pattern Recognition (CVPR), IEEE},
  year = {2017}
}

License

This code is released under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (please refer to LICENSE.txt for details).

Contact:

If you have any questions, please email Angela Dai at adai@cs.stanford.edu.