Pytorch-3D-R²N²: 3D Recurrent Reconstruction Neural Network

This is a Pytorch implementation of the paper "3D-R2N2: A Unified Approach for Single and Multi-view 3D Object Reconstruction" by Choy et al. Given one or multiple views of an object, the network generates voxelized ( a voxel is the 3D equivalent of a pixel) reconstruction of the object in 3D.
See chrischoy/3D-R2N2 for the original paper author's implementation in Theano, as well as overview of the method.

Pre-trained model

For now, only the non-residual LSTM-based architecture with neighboring recurrent unit connection is implemented. It is called 3D-LSTM-3 in the paper.
A pre-trained model based on this architecture can be downloaded from here. It obtains the following result on the ShapeNet rendered images test dataset:

IoU	Loss
0.591	0.093

Installation

The code was tested with Python 3.6.

Download the repository

git clone https://github.com/alexgo1/pytorch-3d-r2n2.git

Install the requirements

pip install -r requirements.txt

Training the network

Download and extract the ShapeNet rendered images dataset:

mkdir ShapeNet/
wget http://cvgl.stanford.edu/data2/ShapeNetRendering.tgz
wget http://cvgl.stanford.edu/data2/ShapeNetVox32.tgz
tar -xzf ShapeNetRendering.tgz -C ShapeNet/
tar -xzf ShapeNetVox32.tgz -C ShapeNet/

Rename the config.ini.example config template file to e.g your_config.ini, and change parameters in it as required.
Run python train.py --cfg=your_config.ini. Or simply python train.py if you named your config file config.ini.

Test your trained model

Run python test.py --cfg=your_config.ini. Or simply python test.py if your config file is named config.ini.
This can be the same config file used for training the model. Note that when testing, you probably want to set resume_epoch to the number of epochs that your model was trained for.

License

MIT License