Comparing Image Representations for Training a Convolutional Neural Network to Classify Gender

In this work, we evaluated the effect of different image representations on the classification performance of a convolutional neural network. Several different methods for normalization of the input data were also considered. The network was discriminatively trained for the task of gender classification of pedestrians. A publicly available dataset was used for training, containing both frontal and rear views of pedestrians. The best result was obtained using grayscale representation as compared to RGB and YUV, giving cross-validated accuracy of 81.5% on the dataset. The performance of the convolutional neural network is competitive and comparable to previous works on the same dataset.