Awesome

FARSE-CNN: Fully Asynchronous, Recurrent and Sparse Event-Based CNN

Official repository of our ECCV 2024 paper "FARSE-CNN: Fully Asynchronous, Recurrent and Sparse Event-Based CNN", by Riccardo Santambrogio, Marco Cannici and Matteo Matteucci.

Installation

This code was developed and tested in a Docker container using the provided Dockerfile. If you do not use Docker, make sure to install the necessary packages in your environment:

pip install -r requirements.txt
pip install torch-scatter --no-index -f https://data.pyg.org/whl/torch-1.13.0+cu116.html

Optionally, install Neptune for logging experiments.

pip install neptune

Data Preparation

Please download (any of) the N-Cars, N-Caltech101 and Gen1 Automotive datasets from their official websites. The automatic download of DVS128 Gesture is supported by tonic.

All datasets should be placed under ./data .

Train

Training settings are controlled using .yaml configuration files. You can find our sample configurations in ./configs. To start training, run the following script:

python train.py --train_cfg={path_to_cfg}

Test

To test a model on object recognition or gesture recognition datasets, use the test.py script. As with training, settings are controlled using .yaml configuration files.

python test.py --test_cfg={path_to_cfg}

To test object detection models, we use the tool provided in Object-Detection-Metrics. You can generate the annotation files running:

python test_detect.py --test_cfg=configs/test_cfg_detect.yaml

Citation

If you find this code useful, please cite:

@inproceedings{santambrogio2024farsecnn,
    author = {Santambrogio, Riccardo and Cannici, Marco and Matteucci, Matteo},
    title = {FARSE-CNN: Fully Asynchronous, Recurrent and Sparse Event-Based CNN},
    booktitle = {The European Conference on Computer Vision (ECCV)},
    year = {2024}
}