A novel supervised approach to learning efficient kernel descriptors for high accuracy object recognition

Discriminative patch-level features are essential for achieving good performance in many computer vision tasks. Recently, unsupervised learning approaches have been employed to design such features based on the similarities of image patches. These approaches, such as kernel descriptors (KD) and efficient kernel descriptors (EKD), have shown superior performance than pre-defined image features (e.g., SIFT or HoG) in object recognition. They gave a kernel generalization of orientation histograms and suggested a promising way to 'grow-up' features based on available information.A major limitation of these approaches is patch similarities are not directly linked to object categories. Therefore, a supervised approach to learning patch-level features that takes into account image class labels is in urgent need. In this paper, we achieve this goal by proposing supervised efficient kernel descriptors (SEKD), in which incomplete Cholesky decomposition is performed jointly with image class label in feature learning. Experimental results on several well-known image classification benchmarks suggest that SEKDs are more compact and have superior discriminative power than previous unsupervised feature descriptors.