ICCV Daily 2021 - Thursday

very small signal is transmitted to the brain. In the event-based camera we try to do the same. Imagine this matrix of pixels and each pixel triggers an event, so in the output you have this streaming of events instead of frames. ” There are three main advantages to this approach. One is that there is sparse data , rather than full images of frames, so it really focuses on what is changing in the scene. Second, it can detect changes , like very fast motion, at microsecond precision. Third, when light changes in the scene, since there is no frame, it can detect and modify very quickly to the change of illumination . “ If you were to imagine reinventing a camera, not for people to look at pretty pictures, but purely for computer vision purposes, would those cameras look like the everyday cameras we all use? ” Keigo poses. “ Probably not. In computer vision, you want to be able to directly access the type of information that you want to know and reason about the scene. With event detection cameras we are reimagining what cameras are supposed to be like. We want to be able to extract information that normal cameras can’t, and speed and dynamic range are important for that. Also, synchronized frames are no longer needed if you can work in a synchronous world and process the data at the speed of the information that’s coming in. ” The event-based camera is asynchronous, but the state of the art is synchronous. This is a fundamentally different way of looking at cameras and images. The majority of the techniques that people have developed for computer vision and image processing fall apart because there is no notion of a picture or a frame, rather it is a continuous processing of information as it comes. It is easy for people to get confused by putting their previous experiences with conventional cameras into the context of this one. There are fundamental ways that you can exploit this camera, but you have to forego some of the prior biases of how you look at images. 11 DAILY ICCV Thursday Introduction to Event Detection Cameras “Each pixel is already making a very small but intentional decision on how it treats the information as it comes in…”