What Models do we Need in Computer Vision? From Optical Flow to Scene Representations
Eddy Ilg
University of Freiburg, Germany
CIS@MPG Colloquium
Eddy holds Master degrees from the University of Southern California in artificial intelligence and from the University of Freiburg in robotics and computer vision. He did his PhD under Thomas Brox and is known for his work on estimating optical flow with convolutional neural networks. Currently, Eddy is a senior research scientist in industry in the domain of augmented reality working on 3D reconstruction and neural scene representations with a focus on object reconstruction in the wild.
Deep learning today is successful in almost any domain of computer vision. The talk will revisit the seminal work of FlowNet to show how deep learning was applied to optical flow and led to a paradigm shift in this domain. Optical flow, disparity, motion and depth boundaries as well as uncertainty estimation with multi-hypothesis networks will be covered and it will be discussed how deep learned models could surpass traditional methods. Asking the more fundamental question what models we need in computer vision, the talk will then progress to recent deep-learned scene representation approaches such as the ones obtained by learned signed distance functions and NeRF and provide a perspective on how computer vision might change in the future.
--
Please contact the MPI-SWS office team for link information.