Code base to reproduce results in :
@article{Bonazzi2025CVPRW,
author = {Bonazzi, Pietro and Vogt, Christian and Jost, Michael and Khacef, Lyes and Paredes-Valles, Federico and Magno, Michele},
title = {Towards Low-Latency Event-based Obstacle Avoidance on a FPGA-Drone.},
journal = {IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop (CVPRW)},
month = {June},
year = {2025},
}
@article{Bonazzi2025IJCNN,
author = {Bonazzi, Pietro and Vogt, Christian and Jost, Michael and Qin, Haotong and Khacef, Lyes and Paredes-Valles, Federico and Magno, Michele},
title = {Fusing Events and Frames with Self-Attention Network for Ball Collision Detection.},
journal = {International Joint Conference on Neural Networks (IJCNN)},
month = {June},
year = {2025},
}
Leave a star to support our open source initiative!⭐️
The name of the main folder with all the modules is eva which stands for Event Vision for Action prediction.
To download the dataset click here.
Click on this link to download the models.
conda create -n eva-env python=3.10 -y
conda activate eva-env
pip install h5py numpy matplotlib opencv-python pandas scipy dtaidistance \
pytorch-lightning torchvision torch fire wandb torchprofile onnx scikit-learn dotenv \
pybind11 tensorboard tensorboardX
Replace the path variable of the dataset (ABCD_DATA_PATH) to your installation directory. Build the docker container:
docker build -t eva-pytorch .
Run the docker container:
docker run --gpus all -it --rm \
-v `ABCD_DATA_PATH`:/workspace/data \
-e DISPLAY=$DISPLAY \
--network host \
--volume /tmp/.X11-unix:/tmp/.X11-unix \
eva-pytorch
In your shell, inside the code directory, install OpenEB, following these instructions :
git clone https://github.com/prophesee-ai/openeb.git --branch 4.6.0
cd openeb && mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTING=OFF
cmake --build . --config Release -- -j $(nproc)
. utils/scripts/setup_env.sh
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib
export HDF5_PLUGIN_PATH=$HDF5_PLUGIN_PATH:/usr/local/lib/hdf5/plugin
export HDF5_PLUGIN_PATH=$HDF5_PLUGIN_PATH:/usr/local/hdf5/lib/plugin
cd ../..
First, visualize a recording 👀
python3 -m eva.data.subclasses.flight --id=1
Then, evaluate the EVS and RGB models 📚 For example Tab. 3 [1] can be reproduced as follows:
python3 -m scripts.eval.cvprw.tab_3
To train the fusion model:
python3 -m scripts.train --inputs_list=["dvs","rgb"] --name=fusion-model --frequency="rgb"
To train the event-based model:
python3 -m scripts.train --inputs_list=["dvs"] --name=dvs-model --frequency="rgb"
To train the rgb-based model:
python3 -m scripts.train --inputs_list=["rgb"] --name=rgb-model --frequency="rgb"
Deploy instructions coming soon.