Many high-level computer vision algorithms suffer in the presence of occlusions caused by multiple objects overlapping in a view. Occlusions remove the direct correspondence between visible areas of objects and the objects themselves by introducing ambiguity in the interpretation of the shape of the occluded object. Ignoring this ambiguity allows the perceived geometry of overlapping objects to be deformed or even fractured. Supplementing the raw image data with a vectorized structural representation which predicts object completions could stabilize high-level algorithms which currently disregard occlusions. Studies in the neuroscience community indicate that the feature points located at the intersection of junctions may be used by the human visual system to produce these completions. Geiger, Pao, and Rubin have successfully used these features in a purely rasterized setting to complete objects in a fashion similar to what is demonstrated by human perception. This work proposes using these features in a vectorized approach to solving the mid-level computer vision problem of object stitching. A system has been implemented which is able extract L and T-junctions directly from the edges of an image using scale-space and robust statistical techniques. The system is sensitive enough to be able to isolate the corners on polygons with 24 sides or more, provided sufficient image resolution is available. Areas of promising development have been identified and several directions for further research are proposed.
Library of Congress Subject Headings
Computer vision; Image processing--Digital techniques; Visual perception; Algorithms
Department, Program, or Center
Computer Engineering (KGCOE)
Kelso, Carl Ryan, "Direct occlusion handling for high level image processing algorithms" (2009). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus