Bag-of-words models playa central role in modern object categorization methods, with intra-category appearance variations commonly accommodated with loose matching thresholds. This solution, however, presents conflicting demands. Recent approaches seek to increase part generalization while avoiding ambiguity by considering co-occurrences and configuration information of pairs or groups of local features. However, learning discriminative compound appearance/configuration features automatically from training data carries a high computational cost.;Empirical evaluation assesses the algorithm's performance using second order spatial features, due to their low demands in terms of training data. The computational cost of the algorithm is modest: discriminative spatial pair features are identified in a fraction of the time required to extract local features. Identified features are shown to provide additional discriminative information: high correlation of the features to the target object classes is maintained in the test set, and their inclusion in a classification system boosts classification performance. A 13.9% gain in the number of images classified correctly is realized on the standard Caltech-256 benchmark dataset, as well as significant improvement in the MSRC Weakly Labeled and v2 datasets.;This dissertation proposes an efficient algorithm for learning such features automatically for object categorization from unsegmented and potentially cluttered training data. Compound features composed of groups of local appearance features detected by means of a low-level interest operator, with optional associated invariant geometric relations create a combined appearance/shape representation with a normalized object-centered coordinate frame. Features are examined for their discriminative value with respect to the target object classes and progressively complex features are built from combinations of simpler features that are highly correlated to the target classes.
展开▼
