Mobile-LRPose: Low-Resolution Representation Learning for Human Pose Estimation in Mobile Devices

Human pose estimation has made great progress in performance due to the development of deep learning. Current methods, including some lightweight networks, usually generate high-resolution heatmaps with rich position information to ensure high accuracy, however, the computational cost is heavy and sometimes unacceptable to mobile devices. In this paper, we construct a network backbone based on the modified MobileNetV2 to only generate low-resolution representations. Then, to enhance the capability of keypoints localization for our model, we also make crucial improvements consisting of bottleneck atrous spatial pyramid, local-space attention, coordinate attention and position embedding. In addition, we design two different network heads for 2D and 3D pose estimation to explore the extensibility of the backbone. Our model achieves superior performance to state-of-the-art lightweight 2D pose estimation models on both COCO and MPII datasets, which achieves 25+ FPS on HUWEI Kirin 9000 and outperforms MoveNet in the same device. Our 3D model also makes nearly 50% and 90% reduction on parameters and FLOPs compared to lightweight alternatives. Code is available at: https://github.com/NanXinyu/Mobile_LRPose.git .