Computer Vision News - June 2023

10 CVPR Best Paper Award Candidate rates, bandwidth requirements, and power consumption. “ Event cameras have been promising, especially for applications like feature tracking, where the benefits of high speed and high dynamic range have been promised , and our work unlocks that in a data-driven fashion, ” Carter ponders. “ At a higher level, our proposed algorithm is quite simple! It’s a straightforward idea that works well, so that’s also something in our favor. ” Thinking about the next steps for this work, the team plan to revisit the initial motivation to use the feature tracker for pose estimation. They want to integrate it with a VO pipeline to improve the overall estimationof thecamerapose. Additionally, an open question remains regarding which features to track. “ In this paper, we predetermined which features we wanted to track, ” Mathias reveals. “ You can imagine that certain features are easier to track in certain scenarios than others, but at the moment, this is completely handcrafted. Eventually, the process could be optimized end to end with a visual odometry or visual-inertial odometry pipeline . ” promising for the application of event cameras , but it’s never been robustly shown to work to such a degree that people think you could apply it directly and gain huge benefits,” Mathias points out. “One issue was that modeling event cameras is very difficult. Even for our lab, which is quite expert in this, it’s not straightforward. In our solution, we overcome the longstanding challenge of event-based feature tracking in a data-driven fashion, letting the benefits of event cameras shine. This is reflected also in the experimental results, where we show over 100% gain over the baseline . It’s not very common in computer vision to have such strong results. ” Nico says the future application of the methodhas likelyplayedarole inthepaper’s success. It demonstrates the possibility of combining traditional image approaches with events, reducing bandwidth in certain applications. While not explicitly shown in the paper, this reduced bandwidth is significant for virtual and augmented reality scenarios . Tracking images with events makes it possible to make real- time decisions about feature track quality and request new images when necessary. This approach can result in lower frame