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.

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
Trained models: trained_models.zip (700mb)

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}
}

Contact:

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

lioqio / cnncomplete