SplitGlass: A Splitting Based Deep Network for Efficient Human Pose Estimation

Recent trends of performing all the image-related tasks on mobile devices results in developing a scalable and efficient model. The current state-of-the-art(SOTA) methods are able to achieve high accuracy by using deep neural networks. However such performance is achieved by compromising the computational complexity. To overcome this, we propose a novel deep network that aims to find the perfect equilibrium between accuracy and efficiency. We use the most common benchmark, the hourglass network as our backbone and propose a splitting-based network in which individual modules within the network focuses on estimating the keypoints for particular regions of the body such as the upper body, lower body, leg, hand, etc. The splitting is done in a hierarchical manner based on the different types of semantic features at each stage. The network with this type of splitting needs a lesser number of parameters to achieve the task of human pose estimation and thereby reducing the computational complexity. The approach to increase efficiency using semantic features based splitting is more generic and can be applied to any deep network. The results of our proposed network is evaluated on the MPII human pose dataset and it achieves a significant reduction in the network parameters as compared to the current SOTA models with similar accuracy.