CVPR Daily - Thursday

high-speed motion and maneuvers, with robust feature tracks to compute the pose using SLAM and VO backends. “ Robust feature tracking is often called the front end of visual odometry pipelines, ” Mathias tells us. “ These VO pipelines are the foundation of mobile robotics because they’re required for control algorithms in the robots to tell the robot where it needs to go or what it needs to do. If you don’t have access to these visual feature tracks, it’s almost impossible to tell the robot what it needs to do from its own perception. ” Previous work in this domain has predominantly relied on standard image cameras, such as video cameras, which have some drawbacks. Existing trackers based on standard camera images are affected by issues such as motion blur in high-speed scenarios, resulting in a loss of scene structure. Additionally, the frame rate of standard image cameras is typically limited to around 20 fps. This work proposes the inclusion of an event camera alongside a traditional camera, which offers higher temporal resolution and enhances the overall performance of feature tracking. “ What is new about this paper is that we have an end-to-end solution to feature tracking that combines both frame-based and event-based domains ,” Mathias explains. “One of the big challenges, if you use both modalities, is how can you associate information from one modality to the other while not losing the advantages of each modality? ” Existing event-based trackers use model assumptions to solve the problem. In this work, the team uses a novel data-driven approach so 4 DAILY CVPR Thursday Best Paper Award Candidate