SyncSpecCnn

Synchronized Spectral CNN for 3D Shape Segmentation.

Introduction

This work is based on our arXiv tech report. In this repository, we release code, data for training Synchronized Spectral CNN for 3D Shape Segmentation. The data we use is from A Scalable Active Framework for Region Annotation in 3D Shape Collections, with a slight re-formatting for our training/test purpose. And the training/test split of the data comes from ShapeNet.

Citation

If you find our work useful in your research, please consider citing:

@article{yi2016syncspeccnn,
  title={SyncSpecCNN: Synchronized Spectral CNN for 3D Shape Segmentation},
  author={Yi, Li and Su, Hao and Guo, Xingwen and Guibas, Leonidas},
  journal={arXiv preprint arXiv:1612.00606},
  year={2016}
}

If you use the data provided, please also considering citing:

@article{yi2016scalable,
  title={A scalable active framework for region annotation in 3d shape collections},
  author={Yi, Li and Kim, Vladimir G and Ceylan, Duygu and Shen, I and Yan, Mengyan and Su, Hao and Lu, ARCewu and Huang, Qixing and Sheffer, Alla and Guibas, Leonidas and others},
  journal={ACM Transactions on Graphics (TOG)},
  volume={35},
  number={6},
  pages={210},
  year={2016},
  publisher={ACM}
}
@article{chang2015shapenet,
  title={Shapenet: An information-rich 3d model repository},
  author={Chang, Angel X and Funkhouser, Thomas and Guibas, Leonidas and Hanrahan, Pat and Huang, Qixing and Li, Zimo and Savarese, Silvio and Savva, Manolis and Song, Shuran and Su, Hao and others},
  journal={arXiv preprint arXiv:1512.03012},
  year={2015}
}

Installation

Install Torch7.

Note that cuDNN and GPU are highly suggested for speed reason. You also need to install a few torch packages (if you haven't done so) including cunn, torchx, optim, matio.

Usage

Fetch data including point cloud sampled from ShapeNet shapes, point features and segmentation labels:
```
 bash getdata.sh
```

These data has been split into different categories and is also split into training/test/validation sets for each category. The data file size is 5GB in total.

Compute Laplacian basis for individual shapes and compute joint Laplacian basis for each shape category:
```
 Matlab/data_preprocessing.m
```
You will need matlab to preprocess the data. There is one sample category having been pre-processed already called Sample, which could be directly used for training.
Train SyncSpecCNN for each category. To see HELP for training script:
```
 cd Lua
 th main.lua -h
```
An example training command is as below:
```
 cd Lua
 th main.lua -s Sample -i 33 -o 4 -ntr 3 -nte 1 -nval 1 -e_b1 20 -e 20 -g 0
```
The segmentation score will be printed as training goes.

Results

Please refer to Table 2 in our arXiv tech report for segmentation IoUs.

License

Our code and data are released under MIT License (see LICENSE file for details).

TODO

Example code for point cloud part label inference.

ericyi/SyncSpecCNN