ICCV Daily 2021 - Wednesday

Previous state-of-the-art methods for 3D hand pose estimation relied on using RGB or depth cameras . However, these cameras have limitations, such as motion blur when there are fast hand motions, and they do not cope very well with night-time recordings. With event cameras , these problems can be solved. In contrast to conventional cameras, which record signals synchronously , event cameras record asynchronous changes in the brightness. “ This is a continuous stream of per pixel brightness changes, and the temporal resolution of events can be up to one microsecond, ” Vlad explains. “ That means the resolution of the equivalent frame rate would be 1 million frames per second. The main challenge is that conventional computer vision methods can’t be applied to this new data modality. New technology has to be built and new methods need to be proposed to solve the problem of hand pose estimation from such asynchronous event streams. ” To solve this, Viktor and Vlad set a milestone to do hand reconstruction at 1000 Hz, which is much faster than before. At this kind of resolution, they knew that if they wanted to apply learning-based techniques, they had to have a dataset that allows it, and a network that allows such fast interference. Viktor Rudnev is a PhD student at the Max Planck Institute for Informatics (MPI-INF). Vladislav Golyanik is Research Group Leader at MPI-INF, leading the 4D and Quantum Vision group. Their work on estimating 3D hand pose from an event stream has been accepted as a poster presentation and they speak to us ahead of their live Q&A session today. EventHands: Real-Time Neural 3D Hand Pose Estimation from an Event Stream 14 DAILY ICCV Wednesday Poster Presentation Viktor Rudnev and Vladislav Golyanik “ With event cameras, these problems can be solved .”