On the use of convolutional neural networks for graphical model-based human pose estimation

Abstract

The recent application of Convolutional Neural Networks (CNNs) to Human Pose Estimation (HPE) from static images have improved estimation accuracy compared to traditional HPE approaches. In particular, a recent novel HPE approach combines a traditional graphical model with CNNs to result in state-of-the-art HPE accuracy, improving the estimation accuracy compared to using either approach alone. However, the accuracy of the CNN used in the hybrid model has not yet been explored, and this paper evaluates the use of CNNs in the hybrid model through investigating different network configurations and fine-tuning the network using pre-trained weights obtained from a large labeled dataset. The proposed CNN configurations not only improve the accuracy of the existing network by up to 2% but also uses fewer parameters, resulting in a higher HPE accuracy and simpler network structure.