Empowering Efficient Human Pose Estimation with Semantic Splitting
Harsh A. Patel, Dhaval K. Patel, Kashish D. Shah, Manob Jyoti Saikia, Bryan J. Ranger · Research Square · 2023
Abstract For real-world applications of human pose estimation (HPE), high efficiency and accuracy are essential.To achieve high accuracy, many state-of-the-art methods employ deep neural networks. However, these approachesoften utilize complex architectures, and performance may be compromised due to high levels of computationalcomplexity. To address this limitation, we propose a novel semantic splitting-based approach that aims to find amore optimal balance between accuracy and computational cost. We incorporate our approach into two existingbenchmarks: the stacked hourglass network, and the simple baseline network. Individual modules within ourproposed network estimate key points for particular body regions such as the upper body and lower body, leg, andhand. The hierarchical splitting of the network is based on different semantic features at each network stage. Thenetwork with this type of splitting needs fewer parameters to accomplish the HPE task, reducing the computationalcomplexity. We evaluated our proposed networks using two benchmark data sets: the MPII human pose data setand the 2017 Microsoft COCO key-point data set. Results demonstrate similar accuracy and a significant reductionin the number of network parameters compared to other state-of-the-art models. Our network is a lightweight andefficient approach for HPE that maintains a high accuracy at decreased computational cost, and has the potentialto be applied to other applications that use deep neural networks.