Senior Principal Research Scientist
Dr. Dong Tian joined InterDigital recently as a Senior Principal Research Scientist after working with Computer Vision Group of Mitsubishi Electric Research Laboratories (MERL) at Cambridge, MA, U.S. between 2010-2018. Previously, he conducted researches with Thomson Corporate Research at Princeton, NJ from 2006 and Tampere University of Technology in Finland from 2002. He has been actively contributing to video related standards in MPEG within projects including H.264/MPEG AVC (2002-), MVC (2006-), H.265/HEVC extensions such as MV-HEVC, 3DV (2010-), and point cloud compression (2016-). His current research interests cover point cloud processing, graph signal processing, deep learning, and computer vision. Besides numerous publications on top-tier conferences and transactions, he holds over 30 US-granted patents. In addition, he serves as technical committee member in multiple IEEE societies. Dong received M.Eng. and B.Eng. degrees on automation from the University of Science and Technology of China (USTC) in 1998 and 1995, respectively; and he was granted Ph.D. at Beijing University of Technology in 2002. He is a senior member of IEEE.
Point cloud data has been emerging and popular for various applications including virtual reality (VR), augmented reality (AR), plus many other computer vision problems. Comparing to conventional image/video data, several fundamental challenges need to be addressed, e.g., its large data rate, irregular sampling structure, varying geometric as well as associated attributes. In this talk, we first review recent progress in point cloud compression and an ongoing MPEG standardization attempt. A framework to conduct fast point cloud resampling is then presented based graph signal processing, which could be utilized to scale down the complexity of many point cloud processing tasks. Last, deep neural networks (DNNs) are generalized for native supporting on point clouds as a new type of signal. For one example, a folding-based autoencoder (AE) architecture FoldingNet is introduced with great potentials for point cloud classification tasks. In short, fundamental techniques are highlighted how point clouds could be successful with more efficient compression, processing and understanding.