IEEE Dayton Section

Abstract: A primary bottleneck in video processing is the readout of large sensor arrays. Typical video contains highly correlated information, which goes unexploited in traditional imaging devices. This talk introduces event cameras, which are a revolutionary hardware design that eliminates the need for large data handling and bypasses the readout of sparse information in large arrays. Furthermore, this talk demonstrates a novel representation for event cameras called TORE volumes. Event representations are utilized to leverage current AI/ML methods for unique event camera data, and optimal encoding methods are an active area of research. TORE volumes have several benefits over other methods (e.g. prioritized encoding, low computational cost, and temporal consistency), making them an ideal replacement for any machine learning solution that struggles to encode sparse event data into a meaningful dense tensor. TORE volumes are evaluated using several public datasets and achieve state-of-the-art performance for human pose estimation, image reconstruction, event denoising, and classification.Speaker(s): Dr. Wes Baldwin, Dayton, Ohio, United States, Virtual: https://events.vtools.ieee.org/m/295624