Improving object recognition accuracy and speed through nonuniform sampling

Silhouette-based shape retrieval and recognition have been well studied, because silhouettes are compact representations of object shape, and because they can be reliably extracted in controlled-environment applications such as digitizing museum collections. In past work, we developed a fast and accurate method for retrieval and recognition of object silhouettes and other closed planar contours. The method is based on a combination of alignment, correspondence, eigenspace dimensionality reduction, and example-based retrieval. Its efficiency and accuracy result from the particular forms of each of these components and the way they are combined. This paper presents two improvements to the method: non-uniform sampling and a new similarity measure. The improved method ranks first in retrieval accuracy in comparison with eight prior methods tested on a benchmark database of 1,400 shapes. Its classification accuracy is 96.8% for the first-ranked class hypothesis, and it returns the correct classification in the top ten hypotheses 99.8% of the time. Average time for retrieval and recognition is approximately 0.6 seconds in Matlab on a PC.