AI & Fundamentals
Domes to Drones: Self-Supervised Active Triangulation for 3D Human Pose Reconstruction - Cristian Sminchisescu, Professor, Lund University / Google Zurich
DATE: Mon, December 9, 2019 - 10:00 am
LOCATION: ICCS - X836, ICICS Computer Science, 2366 Main Mall, Vancouver, BC
Existing state-of-the-art estimation systems can detect 2d poses of multiple people in images quite reliably. In contrast, 3d pose estimation from a single image is ill-posed due to occlusion and depth ambiguities. Assuming access to multiple cameras, or given an active system able to position itself to observe the scene from multiple viewpoints, reconstructing 3d pose from 2d measurements becomes well-posed within the framework of standard multi-view geometry. Less clear is what is an informative set of viewpoints for accurate 3d reconstruction, particularly in complex scenes, where people are occluded by others or by scene objects. In order to address the view selection problem in a principled way, we here introduce ACTOR, an active triangulation agent for 3d human pose reconstruction. Our fully trainable agent consists of a 2d pose estimation network (any of which would work) and a deep reinforcement learning-based policy for camera viewpoint selection. The policy predicts observation viewpoints, the number of which varies adaptively depending on scene content, and the associated images are fed to an underlying pose estimator. Importantly, training the policy requires no annotations - given a 2d pose estimator, ACTOR is trained in a self-supervised manner. In extensive evaluations on complex multi-people scenes filmed in a Panoptic dome, under multiple viewpoints, we compare our active triangulation agent to strong multi-view baselines, and show that ACTOR produces significantly more accurate 3d pose reconstructions. We also provide a proof-of-concept experiment indicating the potential of connecting our view selection policy to a physical drone observer. THis is joint work with Erik Gartner and Aleksis Pirinen.
Cristian Sminchisescu is a Research Scientist leading a team at Google, and a Professor at Lund University. He has obtained a doctorate in computer science and applied mathematics with focus on imaging, vision and robotics at INRIA, under an Eiffel excellence fellowship of the French Ministry of Foreign Affairs, and has done postdoctoral research in the Artificial intelligence Laboratory at the University of Toronto. He has held a Professor equivalent title at the Romanian Academy and a Professor rank, status appointment at Toronto, and has advised research at both institutions. During 2004-07, he was a faculty member at the Toyota Technological Institute at the University of Chicago, and later on the Faculty of the Institute for Numerical Simulation in the Mathematics Department at Bonn University. Cristian Sminchisescu regularly serves as an Area Chair for computer vision and machine learning conferences (CVPR, ECCV, ICCV, AAAI) , is a Program Chair for ECCV 2018, and an Associate Editor of IEEE Transactions for Pattern Analysis and Machine Intelligence (PAMI) and the International Journal of Computer Vision (IJCV). Over time, his work has been funded by the US National Science Foundation, the Romanian Science Foundation, the German Science Foundation, the Swedish Science Foundation, the European Commission under a Marie Curie Excellence Grant, and the European Research Council under an ERC Consolidator Grant. Cristian Sminchisescu's research interests are in the area of computer vision (3d human sensing, reconstruction and recognition) and machine learning (optimization and sampling algorithms, kernel methods and deep learning). The visual recognition methodology developed in his group was a winner of the PASCAL VOC object segmentation and labeling challenge during 2009-12, as well as the Reconstruction Meets Recognition Challenge (RMRC) 2013-14. His work on deep learning of graph matching has received the best paper award honorable mention at CVPR 2018.
Assistant Professor Helge Rhodin, Computer Science Department, UBC