Visual Inference Problems: Structured Outputs and Inverse Graphics
Event details
| Date | 14.12.2016 |
| Hour | 10:00 › 11:00 |
| Location | |
| Category | Conferences - Seminars |
Abstract: Images that we perceive or record are the result of interactions between light and physical objects in the world. These interactions are well understood and can accurately be modeled in modern computer graphics engines. This can be understood as simulation processes that encode our understanding of the generative process. While the simulation process is well described and understood, the inverse of this simulator, e.g., to infer physical and semantic properties of scenes, is largely unsolved. My research aims to describe the process of visual inference in a precise mathematical formulation.
In this talk I will give an overview of our recent work in this domain. I will present an informed sampling approach that recovers accurate posterior distributions for visual inference processes. A second work I will present is the extension of convolutional layers of deep convolutional networks. I will present a generalization of the bilateral filter and show how this results in a sparse and high-dimensional filtering operation that can be used in deep CNNs. This has interesting consequences on a range of application domains. I will discuss several applications including material estimation and also present our latest results in the recovery of detailed 3D human pose from static images.
Biography: Dr. Peter Gehler is a research group leader at the Bernstein Center of Computational Neuroscience (BCCN) of the University of Tübingen and the Max Planck Institute for Intelligent Systems. His work lies at the intersection of machine learning and computer vision. His focus is to develop scene representations that allow a detailed understanding of images and videos. This is a structured output prediction problem and leads to challenging inference problems for which he develops new algorithms.
Dr. Gehler studied computer science at the University of Bielefeld and did his PhD work at the Empirical Inference group at the Max Planck Institute for Biological Cybernetics. He was a postdoctoral researcher at ETH Zurich, temporary Professor at TU Darmstadt, and Junior Research Group Leader at the Max Planck Institute for Informatik in Saarbrücken. In 2014 he was the GCPR program chair. He is an associate editor of PAMI, and serves as area chairs for NIPS, ICML, ECCV, and ICCV.
In this talk I will give an overview of our recent work in this domain. I will present an informed sampling approach that recovers accurate posterior distributions for visual inference processes. A second work I will present is the extension of convolutional layers of deep convolutional networks. I will present a generalization of the bilateral filter and show how this results in a sparse and high-dimensional filtering operation that can be used in deep CNNs. This has interesting consequences on a range of application domains. I will discuss several applications including material estimation and also present our latest results in the recovery of detailed 3D human pose from static images.
Biography: Dr. Peter Gehler is a research group leader at the Bernstein Center of Computational Neuroscience (BCCN) of the University of Tübingen and the Max Planck Institute for Intelligent Systems. His work lies at the intersection of machine learning and computer vision. His focus is to develop scene representations that allow a detailed understanding of images and videos. This is a structured output prediction problem and leads to challenging inference problems for which he develops new algorithms.
Dr. Gehler studied computer science at the University of Bielefeld and did his PhD work at the Empirical Inference group at the Max Planck Institute for Biological Cybernetics. He was a postdoctoral researcher at ETH Zurich, temporary Professor at TU Darmstadt, and Junior Research Group Leader at the Max Planck Institute for Informatik in Saarbrücken. In 2014 he was the GCPR program chair. He is an associate editor of PAMI, and serves as area chairs for NIPS, ICML, ECCV, and ICCV.
Practical information
- General public
- Free
- This event is internal
Organizer
- Dr. Olivier Verscheure
Contact
- Dr. Olivier Verscheure