UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Face detection by facets : combined bottom-up and top-down search using compound templates Holst, Glendon Randal


As detection domains increase in size and complexity, new techniques are needed to effectively search the image and feature space. In this thesis, I explore one such approach to object recognition in the domain of face detection. This approach, dubbed compound templates, is compared to a single template approach. The developed system, Facets, provides an implementation of both techniques to enable fair comparison. The compound template technique uses subfeatures and spatial models to represent a compound object (such as a face). From these compound models, hypothesis-based search then combines top-down and bottom-up search processes to localize the search within the image and feature space. Detected subfeatures become evidence for facial hypotheses, which then guide local searches for the remaining subfeatures based upon the expected facial configuration. The compound technique is described and a comparison of the compound templates technique with a single template technique in a mug-shot style face domain is presented. A description of the implementation, along with issues surrounding the compound templates approach is also provided. Attention is paid to performance, including both efficiency and accuracy. The results are complex; but the strengths, weaknesses, and various trade-offs of the two techniques are detailed. The combined bottom-up and top-down approach of compound templates demonstrates a clear advantage over bottom-up only approaches. The compound templates approach also demonstrates better performance for feature sparse images, detection accuracy, domain coverage, and for domains with increasing size.

Item Media

Item Citations and Data


For non-commercial purposes only, such as research, private study and education. Additional conditions apply, see Terms of Use https://open.library.ubc.ca/terms_of_use.

Usage Statistics