Manhattan-Pyramid Distance: A solution to an anomaly in pyramid matching by minimization

In the field of computer vision, pyramid matching by minimization has gained increasing popularity. This paper points out and discusses an inherent anomaly in pyramid matching by minimization that can affect the performance of classification approaches based on this type of matching. As a solution, a new multiresolution measure, called Manhattan-Pyramid Distance (MPD), is proposed. Systematic evaluations are carried out at the task of instance-based object classification on four object image datasets. Results show that MPD improves object classification performance with respect to a standard approach based on pyramid matching by minimization.