Event-Based Object Detection using Graph Neural Networks

Abstract

Event-based object detection is a challenging but promising task, as the nature of sparsity and asynchrony of events is incompatible with state-of-the-art object detection approaches. Conventional deep neural networks do not take advantage of the event camera's high event sampling rate, low power consumption and robustness of brightness changes. Recent works addresses the problem of redundant computations by using a graph representation to model the feature of event streams that the graph representation and graph neural networks for event streams can efficiently extract the meaningful information and reduce the computational complexity. Nevertheless, there is still room for improvement in terms of accuracy and computation efficiency. In this work, we propose a graph-based architecture and a new mechanism for updating the graph, which significantly increases the capacity of graph neural networks while maintaining highly efficient per-event processing. In object detection task, our model achieves higher accuracy and lower FLOPS per event compared to various synchronous/asynchronous methods. To our belief, the framework we proposed is effective and robust, as well as being a significant reduction in the amount of redundant computation.