Seminar Abstract
“Interactive Segmentation of Objects in Images via Minimization of Quadratic Energies on Graphs”
Computer vision, in general, deals with the semantic interpretation of various objects present in an image. A task that is quintessential to computer as well as human vision, for obtaining such interpretations, is that of object segmentation. Essentially, it refers to the problem of finding within an image, the boundaries of certain objects of interest, or alternately, the regions that correspond to each of these objects.
Since images generally contain numerous objects that are further surrounded by clutter, it is often not possible to define a unique segmentation. In other words, the segmentation problem can be ill-posed when working in an unsupervised framework. Interactive algorithms allow the user to label a few pixels as either object or background, thereby making the segmentation problem well posed. Backed by this motivation, we propose to build an interactive system for the segmentation of objects in images. More specifically, our system will allow the user to mark representative scribbles that indicate certain objects of interest in the image, and consequently produce accurate segmentation of the objects of interest.
Our method proceeds by constructing a weighted combinatorial graph, such that each node in the graph corresponds to a pixel in the image. The edge weights of the graph are defined as measures of similarity between the image features of the nodes that they connect. The segmentation problem is then posed as the minimization of quadratic energies defined on this graph, subject to certain constraints defined by the user marked scribbles. We show that this framework finds equivalent constructions in electrical network theory. Such equivalent constructions can be used to generalize existing methods in order to introduce desired properties in the optimization scheme. In this talk, we will present preliminary work on using such equivalent electrical network constructions in order to devise general purpose object segmentation techniques.
