Probabilistic Image Segmentation

Foreground/background (F/B) image segmentation is a fundamental problem in computer vision. We propose a semi-automatic and unsupervised segmentation approach based on a Bayesian graph-based inference scheme. The inference is shown to provide a probabilistic relaxation of NP-hard high order assignment problems, solved via an efficient spectral scheme. We extend the approach to consider multi-scale cues using a Bayesian formulation, and how it can be extended to saliency detection. The proposes schemes achieve state-of-the-art accuracy when applied to contemporary datasets.

* The research was carried out towards the PhD degree at Bar-Ilan University, under the supervision of Prof. Yosi Keller, faculty of engineering.