FN-MNIST
Description
The Fixational Neuromorphic MNIST dataset is an event-based version of the full MNIST benchmark, where bio-inspired fixational drift movements have been used to drive the camera. Training and test data are provided as separate ZIP archives and each data sample (binary file) contained therein can be loaded using the utility function from the given Python file.
Files
Steps to reproduce
Each original MNIST image is converted to its event-based counterpart by means of a custom setup which is mainly composed of a DAVIS-346 neuromorphic camera from iniVation S.p.A., a PTU-E46 pan-tilt actuator from FLIR Commercial Systems Inc., and a BenQ LCD monitor 1920x1080 @ 144Hz. The neuromorphic camera views the static monitor while randomly drifting around with accurately-modeled bio-inspired microscopic movements. A single data sample consists of all events falling in a 34x34 pixel array during 1.3 seconds of recording (corresponding to the duration of the whole motion sequence used). For the conversion process we use a custom Python-based software toolkit that is shared at this link: https://github.com/simo-net/MovNEye. Further details on the data acquisition procedure are available in the related research paper: link to the open-access publication will be made available upon paper acceptance.