Deep Line Reconstruction for Dynamic Vision Sensor

Abstract

Dynamic Vision Sensor (DVS) is a sensor that records changes in brightness of a scene and has robust characteristics in high dynamic range and motion blue screen than frame-based cameras. However, the event, which is the output of DVS, is very noisy and asynchronous, making it difficult to apply traditional computer vision methodology, especially to define meaningful features. In this paper, we propose a learning-based method for line feature extraction that reflects the spatiotemporal characteristics of events. Each event has very sparse information, and there is no labeled data due to its asynchronous characteristics. Therefore, we present supervised learning through a synchronization process with grayscale images taken together. This method used only events as inputs to predict straight lines well and effectively reduce the amount of computation due to excessive events. Also, these extracted lines are suitable for applying traditional computer vision algorithms.