An Open-source Flexible Toolkit for Efficient Event Camera Data Compression and Loading
Yilun Wu1,2 and Guido C. H. E. de Croon1
1 Micro Air Vehicle Laboratory, Delft University of Technology
2 Machine Learning Group, Centrum Wiskunde & Informatica (CWI)
Note
Code coming soon. We are preparing EVKit for public release and will publish the code in this repository very soon. Watch or star the repository to be notified when it is available.
EVKit is an open-source toolkit that removes the data-loading bottleneck of event-based vision research. It covers both recorded and synthetic event data:
- PODCAST (Pixel-Ordered Delta-Coordinates with Aligned Sign and Timestamp), a new event format that stores events in pixel order instead of time order, giving substantially smaller files and much faster decoding than existing formats.
- Fast data loading that decompresses, filters and augments event streams and converts them into frame-like representations (voxel grids, time surfaces, ...) on the fly, on CPUs alone, so intermediate representations no longer need to be precomputed.
- Synthetic events from high-frame-rate video, stored compactly and converted to event representations at load time with a physically grounded RGB-to-event model.
- Dataset conversion utilities and pre-converted popular datasets hosted on Hugging Face.
A paper describing EVKit is forthcoming; citation details will be added here.
EVKit will be released under the MIT License.