A PyTorch library for training BEV style perception models for self driving tasks. This is unaffiliated with the PyTorch project. This currently includes helpful primitives needed to put together a model.

$ pip install git+https://github.com/d4l3k/torchdrive.git

or

$ git clone --recursive https://github.com/d4l3k/torchdrive.git
$ cd torchdrive
$ pip install -e .

I've been documenting the process for this code. Please see my blog at https://fn.lc/post/3d-detr/ for more details.

3D bounding boxes and velocities for dynamic objects such as cars.

Grids of occupancy around the vehicle trained with differential rendering.

Lane line and drivable space trained purely from image space labels.

Per voxel semantic labels for static objects.

The training dataset for this repo has been collected from my car and thus has lots of personally identifying information so I'm not willing to make it public at this time. If you're interested in contributing or collaborating feel free to reach out. I'm happy to test changes on my own hardware and there may be other options too.

If you have any questions or concerns, please reach out to me either by filing an issue or emailing me at [email protected].

This project is a hobby project and done in my free time. This is non-commercial and no profit has been made from this work.

See the LICENSE file for more information. Some files and functions have different licenses and are marked accordingly.