Overfit Neural Networks as a Compact Shape Representation

C++/PyTorch Implementation of the Paper: "Overfit Neural Networks as a Compact Shape Representation" with CUDA renderer

Paper

@misc{davies2020overfit,
      title={Overfit Neural Networks as a Compact Shape Representation}, 
      author={Thomas Davies and Derek Nowrouzezahrai and Alec Jacobson},
      year={2020},
      eprint={2009.09808},
      archivePrefix={arXiv},
      primaryClass={cs.GR}
}

Render of network trained on cube

Dependencies

Sampler

• Eigen >= 3
• TBB >= 4
• CGAL >= 4
• Boost >= 1.48

Renderer

• CUDA >= 9

Renderer Example

• OpenCV >= 3

Python Bindings

• pybind11 >= 1.1

Building

mkdir build && cd build && cmake ..

Build with your preferred compiler. This was built and tested in Visual Studio 2019 x64.

Options

• BUILD_RENDERER: Turns renderer project on or off.
• BUILD_PYTHON: Builds python bindings for point sampler and renderer (if BUILD_RENDERER is ON)

Installing

By default, building the INSTALL project creates the build/workspace directory with all necessary files for examples and training.