/mesh_reconstruction

Code of single-image 3D reconstruction "Neural 3D Mesh Renderer" by H. Kato, Y. Ushiku, and T. Harada.

Primary LanguagePythonMIT LicenseMIT

Single-Image 3D Reconstruction using Neural Renderer

This is the code for 3D reconstruction in the paper Neural 3D Mesh Renderer (CVPR 2018) by Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada.

Related repositories:

Requirements

Please install neural renderer.

# install neural_renderer
git clone https://github.com/hiroharu-kato/neural_renderer.git
cd neural_renderer
python setup.py install --user
# or, sudo python setup.py install

Testing pre-trained models

First, you need download pre-trained models.

bash download_models.sh

Reconstructing shapes

You can reconstruct a 3D model (*.obj) and multi-view images by the following commands.

python mesh_reconstruction/reconstruct.py -d ./data/models -eid singleclass_02691156 -i ./data/examples/airplane_in.png -oi ./data/examples/airplane_out.png -oo ./data/examples/airplane_out.obj
python mesh_reconstruction/reconstruct.py -d ./data/models -eid singleclass_02958343 -i ./data/examples/car_in.png -oi ./data/examples/car_out.png -oo ./data/examples/car_out.obj
python mesh_reconstruction/reconstruct.py -d ./data/models -eid singleclass_03001627 -i ./data/examples/chair_in.png -oi ./data/examples/chair_out.png -oo ./data/examples/chair_out.obj

Input

Output

Evaluating voxel IoU

You can evaluate voxel IoU of a model on test set by the following command.

python mesh_reconstruction/test.py -d ./data/models -eid multiclass

Mean IoU of pre-trained model is 0.5988, which is slightly different from that in the paper (0.6031). This is mainly because of random initialization of networks.

Training

First, you need download datasets rendered using ShapeNet dataset.

bash download_dataset.sh

This dataset is created by render.py using Blender.

You can train models by the following command.

bash train.sh

This produces almost the same models as the pre-trained models. It takes about three days on Tesla P100 GPU.

Citation

@inproceedings{kato2018renderer
    title={Neural 3D Mesh Renderer},
    author={Kato, Hiroharu and Ushiku, Yoshitaka and Harada, Tatsuya},
    booktitle={The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2018}
}