Duration: 55:37

4,241 views

77

On a broad level, computer graphics involves representing 3D information in 2D. Computer vision can be thought of as the inverse problem - inferring 3D information from a projected representation. This talk will discuss two deep learning approaches to 3D human pose estimation and single-view object reconstruction that attempt to learn about solution feasibility while incorporating simple computer graphics techniques to ensure consistency with observations. The first approach optimizes a GAN to produce a parameterization of the feasible solution space, then seeks a solution in that space which is maximally consistent with observations. The follow-up approach is based on combining these optimization steps into a single nested optimization problem.

See more at https://www.microsoft.com/en-us/research/video/multi-level-optimization-approaches-to-computer-vision/