Efficient object detection using convolutional neural network-based hierarchical feature modeling

Abstract

A hierarchical data-driven object detection framework is addressed considering a deep feature hierarchy of object appearances. The performance of many object detectors is degraded due to ambiguities in inter-class appearances and variations in intra-class appearances, but deep features extracted from visual objects show a strong hierarchical clustering property. Deep features were partitioned into unsupervised super-categories at the inter-class level, and augmented categories at the object level, to discover deep feature-driven information. A hierarchical feature model is built using a latent topic model algorithm, assembling a one-versus-all support vector machine at each node to constitute a hierarchical classification ensemble. Extensive experiments show that the proposed method is superior to state-of-the-art techniques using the PASCAL VOC 2007 and VOC 2012 datasets.