Lu Peng ; Jiang Tao ; Li Yining ; Li Xiangtai ; Chen Kai ; Yang Wenming

Abstract

Real-time multi-person pose estimation presents significant challenges inbalancing speed and precision. While two-stage top-down methods slow down asthe number of people in the image increases, existing one-stage methods oftenfail to simultaneously deliver high accuracy and real-time performance. Thispaper introduces RTMO, a one-stage pose estimation framework that seamlesslyintegrates coordinate classification by representing keypoints using dual 1-Dheatmaps within the YOLO architecture, achieving accuracy comparable totop-down methods while maintaining high speed. We propose a dynamic coordinateclassifier and a tailored loss function for heatmap learning, specificallydesigned to address the incompatibilities between coordinate classification anddense prediction models. RTMO outperforms state-of-the-art one-stage poseestimators, achieving 1.1% higher AP on COCO while operating about 9 timesfaster with the same backbone. Our largest model, RTMO-l, attains 74.8% AP onCOCO val2017 and 141 FPS on a single V100 GPU, demonstrating its efficiency andaccuracy. The code and models are available athttps://github.com/open-mmlab/mmpose/tree/main/projects/rtmo.