Towards Optimal VLAD for Visual Recognition

The Vector of Locally Aggregated Descriptor (VLAD) has achieved great success in a broad range of computer vision tasks and shows its great effectiveness in diverse computer vision tasks. However, it remains unclear about the mathematical foundation of the VLAD and the theoretical connection to basic Algorithms is a theoretical contribution to the practical use of VLAD. And to achieve optimal VLAD for visual recognition, we provide a new formulation of VLAD from a new perspective of match kernels to connect with existing important encoding methods, and we have made a thorough investigation of the roles and effects of the two operations widely used in local feature encoding. What's more, our comprehensive study on VLAD will not only enable to improve the performance but also offer a guide for state-of-the-art algorithms based on local features.