Alignment-Free Gender Recognition in the Wild

Gender is possibly the most common facial attribute automatically estimated from images. Achieving robust gender classification “in the wild,” i.e. in images acquired in real settings, is still an open problem. Face pose variations are a major source of classification errors. They are solved using sophisticated face alignment algorithms that are costly computationally. They are also prone to getting stuck in local minima thus providing a poor pose invariance. In this paper we move the alignment problem to the learning stage. The result is an efficient pose-aware classifier with no on-line alignment. Our efficient procedure gets state of the art performance even with facial poses “in the wild.” In our experiments using “The Images of Groups” database we prove that by simultaneously predicting gender and pose we get an increase of about 5% in the performance of a linear state-of-the-art gender classifier.