Asynchronous Event Processing with Local-Shift Graph Convolutional Network

Event cameras are bio-inspired sensors that produce sparse and asynchronous event streams instead of frame-based images at a high-rate. Recent works utilizing graph convolutional networks (GCNs) have achieved remarkable performance in recognition tasks, which model stream as spatio-temporal graph. However, the computational mechanism convolution introduces redundant computation when aggregating neighbor features, limits low-latency nature events. And they perform synchronous inference process, can not achieve fast response to signals. This paper proposes local-shift network (LSNet), utilizes novel operation equipped with local attention component efficient adaptive aggregation features. To improve efficiency pooling feature extraction, we design node-importance based parallel method (NIPooling) for data. Based on calculated importance each node, NIPooling efficiently obtain uniform sampling results parallel, retains diversity streams. Furthermore, achieving signals, an processing procedure is proposed restrict nodes need recompute activations only those affected by new arrival event. Experimental show cost be reduced nearly 9 times through using further efficiency, while state-of-the-art gesture object recognition.