The purpose of this thesis proposal is to research, develop and evaluate novel face and facial feature detection methods. They are used by various applications in biometric recognition, video surveillance or human computer interaction. The most successful methods that have been proposed lately use boosting algorithms with a cascade architecture and features that are fast to compute. Still various challenges like multiple detections, false alarms and real-time constraints remain to be surpassed in a variety of scenarios.
In this proposal we present two methods: (1) an automatic method for clustering and pruning multiple detections and (2) a context-based model used for object localization. In the first method we propose to improve existing methods by following a more principled way, with less heuristics and better robustness to search parameters. We use a modified version of the Adaptive Mean Shift algorithm to cluster detections and an hierarchical model to prune clustered detections. This model uses the score distribution given by the face and facial feature classifiers
(which we consider as the context) to discriminate between false alarms and true detections. The second method will consist of a greedy-based and geometric constrained model that uses the same contextual information to drive the face and facial feature detection. We intend to design the system as to be independent of the object and classifier type and to be robust to the search parameters.