Title: Maximum Margin Vector Correlation Filter

Abstract: Correlation Filters (CFs) are a class of classifiers which are designed for accurate pattern localization. Traditionally CFs have been used with scalar features only, which limits their ability to be used with vector feature representations like Gabor filter banks, SIFT, HOG, etc. In this paper we present a new CF named Maximum Margin Vector Correlation Filter (MMVCF) which extends the traditional CF designs to vector features. MMVCF further combines the generalization capability of large margin based classifiers like Support Vector Machines (SVMs) and the localization properties of CFs for better robustness to outliers. We demonstrate the efficacy of MMVCF for object detection and landmark localization on a variety of databases and demonstrate that MMVCF consistently shows improved pattern localization capability in comparison to SVMs.