GPI Seminar Series: Jordi Pont

Jordi Pont, Object-based supervised evaluation of image segmentation with the Pascal database
Wednesday January 15th, at 11:00, D5-007

Abstract:
In this talk I am going to talk about the object-based supervised evaluation of image segmentation, which allows us to perform experiments in the extensive database of Pascal. I will first show how one can compare a partition to an object via its upper-bound performance and how to compute it mathematically. I then will extend this analysis to hierarchies to prove that these structures are much better suited for object detection than flat segmentations. I will finally apply the analysis to combining multiple hierarchies, showing that diversity significantly improves the achievable quality. I will present extensive experiments comparing many state-of-the-art segmentation techniques