Hierarchical grouping of salient paths

Supervisors: James Elder and Vida Movahedi
Session: Summer 2015

The ultimate goal of Computer Vision is to enable a machine to see and understand an image or scene, at least as well as a human. An important step towards this goal is to partition an image into regions, each corresponding to an object or entity. This is referred to as image segmentation in the computer vision community. Segmentation is an important step towards image understanding and can enhance the performance of many applications such as object detection, object tracking, surveillance, medical imaging, etc. Our framework for Salient Object Segmentation consists of four stages:

1- Finding a set of line segments in the image that represent color or texture discontinuities (edges)

2- Representing the line segments in a sparse graph model

3- Extracting simple closed contours (cycles) in the above graph

4- Ranking the set of closed contours for output

The goal of the proposed project is to research and develop a hierarchical method for extracting object hypotheses (stage 3 in the above framework). Given a set of open paths of arbitrary length in the graph, the problem under consideration is to group them to find plausible closed contours bounding the salient object in the image. This grouping stage is important, since i) the performance of the final product depends on the quality of the hypotheses formed in this stage, and ii) lowering the computational complexity of this stage results in significant speed up of the solution. In particular, working with a senior graduate student or postdoctoral fellow, the successful applicant will:

1. Run our current method for collecting open paths at different lengths.

2. Develop a method for finding intersections of paths at a certain length (or up to a certain length) with lower complexity than simple line intersection methods

3. Analyze the open paths for completeness- given paths of a certain length (or up to a certain length), what is the lowest achievable error of object contour hypotheses that can be formed by them?

4. Develop a method for combining short open paths for obtaining longer ones, or closed contours

5. Compare results with the currently available (and implemented) greedy method.

Required skills:

1. Good programming skills

2. Good math skills

3. Knowledge of MATLAB programming language

For more information about the lab, visit www.elderlab.yorku.ca.