| In this work we have proposed and implemented a system for general object detection in image databases. Our approach is to combine multiple similarity-based retrieval modules in a K Nearest Neighbor style classifier. The combination is based on Dempster's evidence combination formula.; We have used content based image retrieval systems CIRES and SIMPLIcity that work on the query-by-example model. Both modules generate sets of distances among the query image and database images. We interpret these distances to detect the presence of central objects in images. We have performed experiments for classification---given the query image, determine the most probable class label---and verification---given the query image, determine class membership for a particular class.; In certain domains, detection with a single module produced satisfactory precision. The results of evaluation suggest that the combination approach might yield satisfactory precision when both similarity modules perform comparably. |
