Kernel Coding: General Formulation and Special Cases
This work addresses the challenge of entangled class distributions in image feature space for visual recognition, offering a kernelized approach that is incremental over existing coding techniques like bag of words and sparse coding.
The paper tackles the problem of representing images with compact codes for visual recognition by proposing a general formulation for coding in a high-dimensional Hilbert space to improve class separability, and shows that jointly learning the kernel, dictionary, and classifier yields benefits in experimental evaluations on visual recognition tasks.
Representing images by compact codes has proven beneficial for many visual recognition tasks. Most existing techniques, however, perform this coding step directly in image feature space, where the distributions of the different classes are typically entangled. In contrast, here, we study the problem of performing coding in a high-dimensional Hilbert space, where the classes are expected to be more easily separable. To this end, we introduce a general coding formulation that englobes the most popular techniques, such as bag of words, sparse coding and locality-based coding, and show how this formulation and its special cases can be kernelized. Importantly, we address several aspects of learning in our general formulation, such as kernel learning, dictionary learning and supervised kernel coding. Our experimental evaluation on several visual recognition tasks demonstrates the benefits of performing coding in Hilbert space, and in particular of jointly learning the kernel, the dictionary and the classifier.