Article Preview
TopIntroduction
Human gender recognition can be used in a wide range of real-world applications such as video surveillance. In terms of biometric traits, face and gait may be the most important modalities that can be used for gender classification (Ng et al., 2012). Although gender recognition are intensively studied by the previous literatures, most of them are based on a single dataset (Baluja & Yang, 2007; Li et al., 2008; Moghaddam et al., 2002; Shan et al., 2008; Wang et al., 2010). Unlike the human identification systems, gender recognition should be able to be performed across different datasets in real-world scenarios (Ng et al., 2012). Each dataset has its own database bias due to its own unique data collection environments, yet in the context of face/gait-based gender recognition, most of the previous works simply neglect this issue. Although several popular methods like SVM (Moghaddam et al., 2002; Li et al., 2008), AdaBoost (Baluja & Yang,2007; Wang et al., 2010), PCA+LDA (Shan et al., 2008; Chang et al., 2009) can yield high performance on the same dataset (referred to as intra-dataset), they are seldom evaluated in a cross-dataset manner. In this work, we test these algorithms to see whether they are robust enough against the bias from different datasets. Figure 1 demonstrates several face images from 5 different face datasets while Figure 2 provides some Gait Energy Images (GEI, i.e., average gait silhouette over a gait cycle (Han & Bhanu, 2006)) from 2 different gait datasets. Can you tell the bias pattern for each group of face/gait images in Figures 1 and 2.
Figure 1. Cropped images from the face datasets: (a) AR; (b) FERET; (c) FRGC; (d) LFW; and (e) TFWM
Figure 2. GEI samples: (a) female samples from CASIA-B dataset; (b) male samples from CASIA-B dataset; (c) female samples from USF dataset; (d) male samples from USF dataset
TopExperimental Setup And Results Analysis
In the previous works, high performance can be achieved when conventional machine learning methods like SVM, AdaBoost, PCA+LDA are used for face/gait-based gender recognition. However, gender is a cue across all datasets and should be independent of specific face/gait dataset. Since each dataset has its own bias (Torralba & Efros, 2011) due to its own unique data collection environments, in this work by testing several popular algorithms in a cross-dataset manner, we aim to evaluate the generalization power of these methods, which are important for practical applications. Correct Classification Rate (CCR) is used to measure the performance.