Object detection using scale specific Boosted parts and a Bayesian combiner

This thesis develops new algorithms for object detection in still images. As a start¬ ing point the particular case of face detection is investigatedseeking improvements which can be generalized to other object types. For improving face detection, one possible way is to search for a better representation that is used within the tradi¬ tional detection paradigms. However, since face detection research has seemingly reached a plateau, a more radical approach is pursued here. The approach of this thesis is to identify, develop and evaluate complementary cues which can then be combined with traditional face detection methods. Of interest are cues which help improve detection accuracy and have so far been overlooked or not evaluated thoroughly. Inspired by the failure analysis of a state-of-the art face detector we explore two different candidate cues. The first cue is human skin. The proposed skin detection algorithm combinesa com¬ prehensive skin color model with shape modeis using mutual information matching. An important result is that the combination of color and shape information out-performs purely color-based approaches. Experiments show that the skin cue is in fact complementary to traditional appearance-based face detectors. As a result combinationsof the two can significantly improve the detection rate or precision. The second cue proposed in this thesis is a face's local context which typically contains a person's upper-body Silhouette. The most remarkable property of the developed local context detector is its robustness to resolution degradations. The detector is shown to yield additional correct face detectionswhich are complemen¬ tary to those of traditional face and skin cues. While the skin cue is rather specific for humans, it turns out that the local context idea allows for a fruitful generalization to other object classes: just as the local con¬ text contributes to the detection at small scales, we can form a set of scale-specific object parts to accommodate a set of different scales. Such scale-dedicated parts, in short scaleparts, cover specific parts of the target object at various positions, extents and resolutions. The proposed scaleparts framework builds on boostedclas-sifier cascades for implementing fast and highly discriminant part detectorsand uses Bayesian Networks to combine their Outputs. Experiments in face and car detection demonstrate that the scaleparts approach is robust with respect to a wide ränge of resolution situations. A side product is the first local-to-global (entire body) detector for pedestrians. From an abstract point of view, the scaleparts framework aims to increase …