Ethnicity, gender and age, a pre-annotated face database

Research community achieved considerable progress in face recognition over the past years. Despite this, present face recognition systems are not yet accurate or robust enough to be fully deployed in under-controlled yet high security environments. A number of works have investigated the impact of face categorization on recognition performance, in order to assess the hypothesis that a preliminary face categorization can be used to contain the search space during identification. Categories are usually related to soft-biometrics, such as gender, age, ethnicity. More features can also be used at the same time to define categories (e.g. gender and age). The underlying assumption is that, during identification operations, a sample image is only matched with those pertaining to the same category. Experimental results demonstrate that face categorization based on important visual characteristics such as gender, ethnicity, and age generally improve recognition accuracy, while reducing operation time. On the other hand, it is difficult to appropriately set up related experiments, since available datasets are not organized according to any categorization. Moreover, it is often the case that some features (e.g. ethnicity or gender) are not uniformly represented. For instance, the ethnicity of the research group gathering a dataset, and therefore the location where the enrollment operations are performed, often influences the prevailing ethnical composition of the dataset. As a further example, since most datasets are gathered by enrolling volunteer students, the prevailing age range in most datasets is 20-35. Our contribution relies in an automatic procedure to build a larger multi-racial database, starting from the most popular among the available ones, which automatically reproduces the ethnicity/gender/age categorization that we manually performed in our lab.