Title: Learning-based 3D understanding of deformable objects using weak and self-supervision
Speaker: Iasonas Kokkinos, Senior Engineering Manager in Snap and Associate Professor in the Department of Computer Science of University College London (UCL).
Abstract: Understanding objects in 3D from a single 2D image is effortless for humans, yet remains challenging for computer vision, largely due to the scarcity of explicit 3D supervision. In this presentation we will discuss how to mitigate this by exploiting 2D datasets through weak- and self-supervision.
We will start with the DensePose, HoloPose and MeshPose line of works for 3D human understanding. We will see how these methods incorporate 3D geometric modeling both in the representation and the training objectives, allowing us to build increasingly simple, efficient, and accurate models. We will cover real-time Augmented Reality applications built with the resulting models for the human body and hands.
We will then turn to the broader problem of understanding generic 3D deformable categories such as horses, cows, dogs or cats. We will see that even in the absence of prior knowledge about the object structure, 3D deformations can be injected into the model representation and used to form self-supervised training objectives for 3D generative models. We will see that the resulting models provide disentangled control of pose, non-rigid motion and appearance in a manner similar to 3D morphable face models - but without having access to any 3D supervision during training.
Bio: Iasonas Kokkinos is Senior Engineering Manager in Snap and Associate Professor in the Department of Computer Science of University College London (UCL).
Iasonas obtained his D.Eng in 2001 and PhD in 2006 from NTUA and his Habilitation (HDR) in 2013 from University Paris-Est. He was a postdoc in UCLA until 2008, and then joined the faculty of Ecole Centrale Paris where he stayed until 2016. In 2016 he joined UCL’s Department of Computer Science as well as Facebook’s AI Research (FAIR) Lab. In 2018 he left FAIR to build Ariel AI, focusing on real-time monocular human reconstruction for Augmented Reality. Following the acquisition of Ariel AI by Snap Inc., he joined Snap full-time as Engineering Manager.
His research interests are focused on 3D human understanding for augmented reality as well as general–purpose computer vision models. His works include the Deeplab, UberNet and DensePose systems that have been impactful in both academia and industry. He publishes, reviews, and regularly serves as Area Chair in all major computer vision conferences (CVPR,ICCV,ECCV).
________________________________________________________________________________
Meeting ID: 394 176 905 976
Passcode: PdJJWw
________________________________________________________________________________
