This keynote presentation explores the application of bilevel optimization to the challenging problem of optical flow estimation in computer vision. We address the hierarchical nature of this task, which involves both reconstruction of the flow field and learning of model parameters. Our approach formulates optical flow computation as a bilevel optimization problem, where the lower-level focuses on flow field reconstruction, and the upper-level learns the optimal regularization parameter.

We demonstrate how this formulation allows us to minimize the discrepancy between estimated and ground truth flow while preserving the benefits of regularization. The presentation will delve into the theoretical aspects of this approach by discussing the existence of solutions and deriving conditions for suitability of the employed regularization. This analysis provides valuable insights into the interplay between reconstruction and learning in optical flow estimation which can offer a robust framework for parameter tuning in complex vision tasks.

El Mostafa Kalmoun received his doctorate in Applied Mathematics (Optimization) from Cadi Ayyad University, Morocco in 2001, then was a recipient of the MIF grant at Niigata University, Japan, and held a postdoctoral position at Erlangen-Nürnberg University, Germany from 2001 to 2004. Afterwards, he joined Cadi Ayyad University as assistant professor and became associate professor in 2007 and full professor in 2015. He was a visiting professor in Spain, jointly at CRM and Pompeu Fabra University in Barcelona in 2007-2008, and later at University Utara Malaysia from 2009 to 2011. Before joining Al Akhawayn University, he was an associate professor at King Khalid University, Saudi Arabia from 2011 to 2018, and Qatar University from 2018 to 2021. He was awarded fellowships and grants from DAAD and DFG in Germany, Spanish Ministry of Education and Sciences, Spanish Agency of International Cooperation, French Cooperation Service, and The World Academy of Sciences. His current research interests involve vector optimization and computational mathematics with applications to image processing, nanocomposites, flow dynamics, etc.