Impact of frame-based representations for event-based data in the field of gesture recognition
Résumé
Gesture recognition using event-based data represents a promising direction in neuromorphic computing, leveraging the asynchronous nature of dynamic vision sensors (DVS) to overcome limitations of traditional frame-based video processing. However, before developing fully spiking solutions, we aim to understand what is the most appropiate use that can be done of event-spike data in presence of a basic convolutional architecture. This work provides insights on the behavior of different conversion methods such as aggregating frames (time-window, spike-count, n-bis), binary representation or time surfaces. These representations are challenged on accuracy metrics in the framework of gesture recognition.