| Object recognition' is difficult due to wide variations in view angles, illumination directions, occlusions, poses, and differences among objects within a class. Besides the problem of learning the object model from data, we need to be able to detect objects efficiently. Since we cannot consider all possible configurations and check them against the image evidence, we must determine which parts of the configuration space should be explored, avoiding its full evaluation. In this work, we propose two classes of methods to group the candidate parts by exploring only a small fraction of the set of all configurations.; Our first class of methods prunes the correspondence search by rejecting small assemblies of candidate parts by determining when it is impossible to augment them with other parts to yield the object of interest. We propose a classifier that determines when a segment assembly corresponds to a person.; Our second grouping method makes the search efficient by recursively decomposing the search space into conditionally independent components, and performs the search in each component and then combines the results. We model the objects using mixtures of trees, which represent the individual part appearances, as well as the relationships between parts (such as the distances and angles between human limbs) and the aspect (which parts are visible and which are absent).; We have applied mixtures of trees to human tracking and view-invariant face detection. Our human tracker is completely automatic; it does not require the user to specify the initial configuration of the person, nor does it rely on background, subtraction. We are able to determine the configuration of the person, in a wide variety of activities and at low frame-rate. The model for a motion sequence consists of a mixture of trees for each frame, augmented with temporal links that enforce consistency between limbs in different frames. Even though these links violate the conditional-independence assumptions that make exact inference on mixtures of trees efficient, we show an algorithm for efficient approximate inference with these links in place. (Abstract shortened by UMI.)... |
