On random hyper-class random forest for visual classification

Random forest is an effective ensemble classifier. It has shown to be efficient with good generalization performance thanks to the randomness in training. However, for visual classification problems, where high dimensions and large number of classes exist, the impurity measurement in training the tree nodes split function not only neglects the strong conditional dependencies among visual attributes, but also leads to rather weak base classifiers, which may not reflect enough discriminative capability. Addressing this, we develop a random hyper-class random forest (RHC-RF) for visual classification tasks in this paper. During training each tree node, the entire class space is randomly partitioned into two hyper-classes and a weak learner targeting at separating the two hyper-classes is trained, where a linear classifier is learned with multiple feature attributes. This novel training scheme well preserves the randomness by random partition of the label space, and at the same time, the objective of training each node directly targets at discriminating classes in a divide-and-conquer-like manner. The proposed method is expected to benefit from the improved discrimination for visual classification compared with conventional random forests while preserving good generalization. Extensive experiments on various multi-class and high-dimensional visual classification tasks (including scene and object classification, image-based food identification, handwritten digit recognition and face recognition) demonstrate the superior accuracy, as well as its compactness and robustness achieved by the proposed method.