On the Generalization Power of Face and Gait in Gender Recognition

Human face/gait-based gender recognition has been intensively studied in the previous literatures, yet most of them are based on the same database. Although nearly perfect gender recognition rates can be achieved in the same face/gait dataset, they assume a closed-world and neglect the problems caused by dataset bias. Real-world human gender recognition system should be dataset-independent, i.e., it can be trained on one face/gait dataset and tested on another. In this paper, the authors test several popular face/gait-based gender recognition algorithms in a cross-dataset manner. The recognition rates decrease significantly and some of them are only slightly better than random guess. These observations suggest that the generalization power of conventional algorithms is less satisfied, and highlight the need for further research on face/gait-based gender recognition for real-world applications.