A Learning Probabilistic Approach for Object Segmentation

This paper proposes a new method for figure-ground image segmentation based on a probabilistic learning approach of the object shape. Historically, segmentation is mostly defined as a data-driven bottom-up process, where pixels are grouped into regions/objects according to objective criteria, such as region homogeneity, etc. In particular, it aims at creating a partition of the image into contiguous, homogenous regions. In the proposed work, we propose to incorporate prior knowledge about the object shape and category to segment the object from the background. The segmentation process is composed of two parts. In the first part, object shape models are built using sets of object fragments. The second part starts by first segmenting an image into homogenous regions using the mean-shift algorithm. Then, several object hypotheses are tested and validated using the different object shape models as supporting information. As an output, our algorithm identifies the object category, position, as well as its optimal segmentation. Experimental results show the capacity of the approach to segment several object categories.