gdbbq / 3d-r2n2-pytorch

PyTorch version of 3D-R2N2 三维重建

3D-R2N2-PyTorch

This is a pytorch version of 3D-R2N2. Original repo: https://github.com/chrischoy/3D-R2N2

Installation

The repo was tested with python3.6, cuda 10.1, pytorch 1.4.0. You can follow the instruction below to install the virtual environment.

• Get the source code.

git clone https://github.com/heromanba/3D-R2N2-PyTorch.git

• Install anaconda(https://docs.anaconda.com/anaconda/install/).

• Create virtual environment and install required packages.

cd 3D-R2N2-PyTorch
conda create -n 3D-R2N2 python=3.6
conda activate 3D-R2N2
pip install -r requirements.txt

Demo

• Download pretrained model(ResidualGRUNet), and put checkpoint.pth under output/ResidualGRUNet/default_model.

  Google drive link(https://drive.google.com/open?id=1LtNhuUQdAeAyIUiuCavofBpjw26Ag6DP)

  Baidu pan link(链接: https://pan.baidu.com/s/12YK4mnQNx9xdCjzV7zx7GA 提取码: 66nf)

• Run The predicted object will be saved to prediction.obj.

python demo.py

Train

Prepare dataset

• Use the same dataset as mentioned in the original repo.

  ShapeNet rendered images http://cvgl.stanford.edu/data2/ShapeNetRendering.tgz

  ShapeNet voxelized models http://cvgl.stanford.edu/data2/ShapeNetVox32.tgz

• Extract data into ShapeNet directory, the file structure in ShapeNet should be like this:

ShapeNet/
    |
    |----ShapeNetRendering/
    |
    |----ShapeNetVox32/
    |

• Change some parameters. You can change parameters in experiments/scripts/res_gru_net.sh or lib/config.py

• Run.

bash experiments/scripts/res_gru_net.sh

License

MIT License