Segmentation Based Structure Matching
Hongzhi Wang, John Oliensis · 2007
This paper investigates how to connect low level image segmentation with high level vision tasks. A segmentation gives a middle-level image rep-resentation which is robust to the intensity and structure variations that cause problems for matching based on low-level descriptors. A segmentation de-scriptor can give more accurate and efficient matching of boundaries than methods based on boundary representations. Our recognition tests on sev-eral databases verify that when segmentations are reliable the segmentation descriptor can provide robust matching results for recognition. To address the problem caused by unreliability of low level segmentation, we propose a Bayesian structure matching technique. Instead of only basing on the “ max-imum likelihood ” segmentation, all segmentations should be evaluated and combined in the Bayesian framework. The resulting matching technique is more reliable and more practical. Furthermore, it offers insights about low level image segmentation as well. We demonstrate successful results in both fields. 1 1