Analysis of gender recognition methods' robustness

A lot of studies are conducted in the area of gender recognition, new methods are constantly proposed for image alignment, feature extraction and classification. But the vast majority of the publications have a certain shortcoming: they only evaluate the methods' performance on a fixed set of images, often of a constant good quality, and pay little or no attention to how it would perform under different, more realistic conditions. The problem of methods robustness is of importance when defining areas of application for known methods and choosing a method to use in a particular situation. This paper aims to encourage more research attention to the problem of robustness and presents a study on several methods' performance in presence of some simple complicating factors such as alignment error or noise. The results show that the method demonstrating the best results on an initial image base proves to be more sensitive to input quality than others and thus loses the competition which proves the importance of such research.