Most discriminative techniques for detecting instances from object categories in still images consist of looping over a partition of a pose space with dedicated binary classifiers. This strategy is inefficient for a complex pose, i.e., for fine-grained descriptions: i) fragmenting the training data, which is inevitable in dealing with high in-class variation, severely reduces accuracy; ii) the computational cost at high pose resolution is prohibitive due to visiting a massive pose partition.
To overcome data-fragmentation I will discuss a novel framework centered on pose-indexed, stationary features, which allows for efficient, one-shot learning of pose-specific classifiers. Such features assign a response to a pair consisting of an image and a pose, and are designed so that the probability distribution of the response is constant if an object is actually present. To avoid expensive scene processing, the classifiers are arranged in a hierarchy based on nested partitions of the pose, which allows for efficient search. The hierarchy is then "folded" for training: all the classifiers at each level are derived from one base predictor learned from all the data. The hierarchy is "unfolded" for testing: parsing a scene amounts to examining increasingly finer object descriptions only when there is sufficient evidence for coarser ones. I will illustrate these ideas by detecting and localizing cats in highly cluttered greyscale scenes. This is joint work with Francois Fleuret.