A paper recently published by Prophesee’s team presents HATS: Histograms of Averaged Time Surfaces for Robust Event-based Object Classification.
The paper was selected to be presented at CVPR 2018, a highly prestigious and selective conference.
There were 2 key motivations behind the paper :
The lack of low-level representations and architectures for event-based sensors
HATS use a new event-based feature representation and a new machine learning architecture for object classification, described in the poster below.
These are designed to take advantage of the higher temporal resolution, and local memory of event-based sensors.
Accuracy using a linear classifier
faster than CSNN
The absence of large real-world event-based datasets
HATS was released along with the first large real-world event-based dataset for object classification: N-CARS.
HATS achieves 90% accuracy for car classification results using a linear classifier, 39x faster than convolutional spiking neural networks – the closest competitor. These results are shown in figures in the poster below