Gender classification using RGB-D videos

Abstract

Gender classification is used in applications as a soft feature or attribute in biometrics to help identify people. Using gender classification as an indexing technique can boost the performance of facial-biometric. If the face images are obtained using high quality camera then only RGB information is sufficient for gender classification. However, in surveillance scenario RGB face images are of low quality and have covariates such as pose, illumination, expression and distance. Therefore, in such scenarios depth information can be utilised to improve the performance of gender classification. Low-cost depth sensors such as Microsoft Kinect provide the depth images along with corresponding RGB color images. These low cost (Kinect) devices can be used for video surveillance; however, not much research has been focused on RGB-D (RGB and Depth data) video data obtained from these devices. In this research, we present a novel gender classification algorithm that extracts features using multiple algorithms from RGB-D videos. While most of the work in gender classification has focused on handcrafted feature extraction techniques such as Uniform Local Binary Pattern and Gradient Local Binary Pattern, we have also studied effectiveness of learned feature extraction techniques such as Stacked denoising autoencoder on gender classification. We also present a score level fusion of handcrafted features and learned features, which significantly improves the performance of gender classification. The proposed algorithm is evaluated on KaspAROV dataset, which contains RGB-D video data obtained from Microsoft Kinect device. This dataset encompasses challenges of varying conditions related to illumination, pose, expression, low image quality and distance. The experiments are also performed on Eurecom Kinect dataset. On both the databases the proposed algorithm achieves state-of-the-art results.