Voxelnet (defunct)

This is a project with the goal of re-implementing VoxelNet (Zhou, Tuzel, 2018).

This is not finished! This is an implementation of Voxelnet without CUDA dependencies. Just Tensorflow 2.0 Keras.

Specifically, the goal is that this project:

1. Implements the novel layers and loss described in the paper,
2. Builds a model from these layers,
3. Trains and evaluates against the relevant dataset,
4. Provides pre-trained weights.

Motivation for this implementation

I had too much difficulty using other codesets and implementations. A lot of them share the same 'lineage' and so I wanted to start a new lineage that I found easier to use.

I wanted this implementation to:

1. Emphasize code-quality,
2. Use pure-Python + Tensorflow, minimizing extra dependencies,
3. Prefer functional approaches
4. Use reasonable default arguments to avoid parameter wrangling,
5. Modularity, so this works more like a "flat library" rather than a "deep system",
6. Parameter transparency: Avoid dicts-as-arguments. (No more params or args arguments!)

So, we want our code to look like this:

Architecture

• The notebooks provides examples and whatnot as I worked through implementation details. This includes unfinished parts and notes.
• The library.py provides every function and Keras layer we use in this project.
• The model.py provides an instantiation of the model using the layers in library.py.

License

This is a WIP, and license is subject to change.