Mobile-friendly and multi-feature aggregation via transformer for human pose estimation

Human pose estimation is pivotal for human-centric visual tasks, yet deploying such models on mobile devices remains challenging due to high parameter counts and computational demands. In this paper, we study Mobile-Friendly and Multi-Feature Aggregation architectural designs for human pose estimation and propose a novel model called MobileMultiPose. Specifically, a lightweight aggregation method, incorporating multi-scale and multi-feature, mitigates redundant shallow semantic extraction and local deep semantic constraints. To efficiently aggregate diverse local and global features, a lightweight transformer module, constructed from a self-attention mechanism with linear complexity, is designed, achieving deep fusion of shallow and deep semantics. Furthermore, a multi-scale loss supervision method is incorporated into the training process to enhance model performance, facilitating the effective fusion of edge information across various scales. Extensive experiments show that the smallest variant of MobileMultiPose outperforms lightweight models (MobileNetv2, ShuffleNetv2, and Small HRNet) by 0.7, 5.4, and 10.1 points, respectively, on the COCO validation set, with fewer parameters and FLOPs. In particular, the largest MobileMultiPose variant achieves an impressive AP score of 72.4 on the COCO test-dev set, notably, its parameters and FLOPs are only 16% and 18% of HRNet-W32, and 7% and 9% of DARK, respectively. We aim to offer novel insights into designing lightweight and efficient feature extraction networks, supporting mobile-friendly model deployment.