Coarse-to-fine classification is an efficient way of organizing object recognition in order to accommodate a large number of possible hypotheses and to systematically exploit shared attributes and the hierarchical nature of the visual world. The basic structure is a nested representation of the space of hypotheses and a corresponding hierarchy of (binary) classifiers. In existing work, the representation is manually crafted. Here we introduce a design principle for recursively learning the representation and the classifiers together. This also unifies previous work on cascades and tree-structured search. The criterion for deciding when a group of hypotheses should be "retested" (a cascade) versus partitioned into smaller groups ("divide-and-conquer") is motivated by recent theoretical work on optimal search strategies. The key concept is the cost-to-power ratio of a classifier. The learned hierarchy consists of both linear cascades and branching segments and outperforms manual ones in experiments on face detection.

We introduce a stochastic model to characterize the online computational process of an object recognition system based on a hierarchy of classifiers. The model is a graphical network for the conditional distribution, under both object and background hypotheses, of the classifiers which are executed during a coarse-to-fine search. A likelihood is then assigned to each history or "trace" of processing. In this way, likelihood ratios provide a measure of confidence for each candidate detection, which markedly improves the selectivity of hierarchical search, as illustrated by pruning many false positives in a face detection experiment. This also leads to a united framework for object detection and tracking. Experiments in tracking faces in image sequences demonstrate invariance to large face movements, partial occlusions, changes in illumination and varying numbers of faces.

We propose merging face detection and face tracking into a single probabilistic framework. The motivation stems from a broader project in algorithmic modeling, centered on the design and analysis of the online computational process in visual recognition. Detection is represented as a tree-structured graphical network in which likelihoods are assigned to each history or "trace" of processing, thereby introducing a new probabilistic component into coarse-to-fine search strategies. When embedded within a temporal Markov framework, the resulting tracking system yields encouraging results.

Our objective is to track multiple faces in a video sequence. We combine a frame-based face detection algorithm based on a coarse-to-fine search with a dynamic Markov model for the joint evolution of poses and observations. The link between the two is the trace model, a probability distribution on the global history of online processing during static face detection. The trace is our frame-based observation and its conditional distribution, given the states of interest (namely multiple face poses and “background”), serves as our likelihood functionP (observation|state). Consequently, the resulting framework unites face detection and face tracking in a single probabilistic setting. Due to continuously updating detections, there are no prior restrictions on face movements. The tracker is tested on real image sequences and accommodates large movements, partial occlusions, varying numbers of faces, and changes in illumination.

Making decisions based on a linear combination L of features is of course very common in pattern recognition. For distinguishing between two hypotheses or classes, the test is of the form sign (L - /spl tau/) for some threshold /spl tau/. Due mainly to fixing /spl tau/, such tests are sensitive to changes in illumination and other variations in imaging conditions. We propose a special case, a "self-normalized linear test" (SNLT), hard-wired to be of the form sign (L/sub 1/ - L/sub 2/) with unit weights. The basic idea is to "normalize" L/sub 1/, which involves the usual discriminating features, by L/sub 2/, which is composed of non-discriminating features. For a rich variety of features (e.g., based directly on intensity differences), SNLTs are largely invariant to illumination and robust to unexpected background variations. Experiments in face detection are promising: they confirm the expected invariances and out-perform some previous results in a hierarchical framework.