Spike context: A neuromorphic descriptor for pattern recognition

Although the neuromorphic vision community has developed useful event-based descriptors in the recent past, a robust general purpose descriptor that can handle scale and rotation variations has been elusive to achieve. This is partly because event cameras do not output frames at fixed intervals (like standard cameras) that are easy to work with, but an asynchronous sequence of spikes at microsecond to millisecond time resolutions. In this paper, we present Spike Context, a spatio-temporal neuromorphic descriptor that is inspired by the distribution of photo-receptors in the primate fovea. To demonstrate the effectiveness of the spike context descriptors, they are employed as semi-local features in the bag-of-features classification framework. In the first set of experiments on the N-MNIST dataset, we obtained very high results compared to existing works. In addition, we outperformed the state-of-the-art algorithms on the smaller MNIST-DVS dataset. Finally, we demonstrate the ability of the descriptor in handling scale variations by using the leave-one-scale-out protocol on the MNIST-DVS dataset.