Rethinking Keypoint Representations: Modeling Keypoints and Poses as Objects for Multi-Person Human Pose Estimation

In keypoint estimation tasks such as human pose estimation, heatmap-based\nregression is the dominant approach despite possessing notable drawbacks:\nheatmaps intrinsically suffer from quantization error and require excessive\ncomputation to generate and post-process. Motivated to find a more efficient\nsolution, we propose to model individual keypoints and sets of spatially\nrelated keypoints (i.e., poses) as objects within a dense single-stage\nanchor-based detection framework. Hence, we call our method KAPAO (pronounced\n"Ka-Pow"), for Keypoints And Poses As Objects. KAPAO is applied to the problem\nof single-stage multi-person human pose estimation by simultaneously detecting\nhuman pose and keypoint objects and fusing the detections to exploit the\nstrengths of both object representations. In experiments, we observe that KAPAO\nis faster and more accurate than previous methods, which suffer greatly from\nheatmap post-processing. The accuracy-speed trade-off is especially favourable\nin the practical setting when not using test-time augmentation. Source code:\nhttps://github.com/wmcnally/kapao.\n