Deep Learning for Vision-based Measurement

点击翻译按钮获取中文摘要

Deep learning techniques are remarkably successful on detection and recognition tasks in computer vision, reaching better than human performance in some specific applications. In my research I will develop novel learning-based algorithms and methods for virtual reality where data driven interaction and environment modelling, as well as user interfaces, can significantly benefit from advances in computer vision. Visual tracking allows different virtual reality and augmented reality devices to remain registered and synchronized. Motion estimation of surfaces, objects and actors is crucial in motion capture and 3D interaction modelling. Multi-view stereo captures real-word environments enabling navigation through these environment based on geometric relationships. My research program will significantly improve visual tracking, motion and stereo algorithms with learning-based techniques by focusing on vision based measurement. Vision based measurement uses the camera as a measurement instrument to obtain a measurand through an associated measurement procedure and with an uncertainty. This is crucial but often overlooked in virtual and augmented reality where different sensors and sensing results need to be fused with each other but also combined with physical reality. The geometry but also the appearance of captured objects and characters must not only appear realistic on their own but also when integrated into the whole virtual or augmented reality. In my research program, I will develop methods that on the one hand use physical constraints on the learning and on the other hand use learning to obtain physically plausible models. I will work on making consistent long-term tracking possible, develop real-time learning methods for motion estimation and 3D capture, thereby advancing the state-of-the-art in virtual and augmented reality through a focus on vision based measurement.