Skip to content

torch models and primitives for birdseye view and voxel models for self driving tasks.

License

Notifications You must be signed in to change notification settings

d4l3k/torchdrive

Repository files navigation

torchdrive

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.

Install from Source

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

or

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

Background

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

3D Object Detection

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

Voxel Occupancy

Grids of occupancy around the vehicle trained with differential rendering.

BEV Lane Lines and Drivable Space

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

Semantic Voxel

Per voxel semantic labels for static objects.

Data Access

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.

Contact

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

License

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.

About

torch models and primitives for birdseye view and voxel models for self driving tasks.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published