Monday, April 29, 2013 - 17:15

Ecole Centrale Paris - Center for Visual Computing

Computer vision as inverse graphics: efficient algorithms for model-based image understanding

The model-based, analysis-by-synthesis approach has served as a rich source of ideas for computer vision. The conceptual appeal of this approach is however marred by the computational complexity of the resulting 'inverse' optimization problems.
In this talk, I will present recent advances on accelerating model-based computer vision. I will start with branch-and-bound optimization for deformable object detection in natural images, demonstrating how simple geometric reasoning can result in multifold acceleration of state-of-the-art object detection systems.
I will then proceed to 'shape grammar parsing', which can be understood as inverse procedural modeling. I will describe how reinforcement learning can be used to perform optimization with grammatical models with structure variability and present applications to the interpretation of building facades.
In the final part of the talk, time permitting, I will outline connections with 3D data modeling and analysis, and address in particular the construction of invariant 3D surface descriptors.