Coordinate regression for event-based data

The repository demonstrates coordinate regression for event-based data with spiking neural networks. Specifically, we contribute:

1. A dataset of event-based vision (EBV) videos for coordinate regression and pose estimation
2. A method for differentiable coordinate transform (DVS) for spiking neural networks
3. Translation-invariant receptive fields that outperforms similar artificial neural network models

Usage

To train the models, follow the below steps

1. Download the dataset via this link and unpack it to a folder you can recall, say /tmp/eventdata.
2. Ensure you have a Python installation with PyTorch and Norse installed.
   • After installing the necessary PyTorch version, you can install the dependencies from the requirements.txt-file by typing: pip install -r requirements.txt
3. Enter the coordinate-regression folder and run the learn_shapes.py file with the dataset directory and model type to start training
   • As an example, run python learn_shapes.py --data_root=/tmp/eventdata --model=snn
     • Four models are available: ann, annsf, snn, and snnrf
     • For training parameter descriptions and help, type python learn_shapes.py --help

Authors and Contact

• Jens E. Pedersen <[email protected]> (Twitter @jensegholm)
• Juan P. Romero B.
• Jörg Conradt

Acknowledgements

This work has been performed at the Neurocomputing Systems Lab at KTH Royal Institute of Technology and funded by the Human Brain Project and the AI Pioneer Centre.

Please cite the work as follows:

@inproceedings{Pedersen_Singhal_Conradt_2023,
  address={New York, NY, USA},
  series={NICE ’23},
  title={Translation and Scale Invariance for Event-Based Object tracking},
  ISBN={978-1-4503-9947-0},
  url={https://dl.acm.org/doi/10.1145/3584954.3584996},
  DOI={10.1145/3584954.3584996},
  booktitle={Proceedings of the 2023 Annual Neuro-Inspired Computational Elements Conference}, publisher={Association for Computing Machinery}, author={Pedersen, Jens Egholm and Singhal, Raghav and Conradt, Jorg},
  year={2023},
  month=apr,
  pages={79–85},
  collection={NICE ’23}
}

License

This work is licensed under LGPLv3.