Fast One-Stage Human Pose Estimation by Polar Coordinate Representation
Jiahua Wu, Hyo Jong Lee · 2022
The typical bottom-up human pose estimation methods leverage two-stage approaches to detect individual instances and predict keypoints, which leads to high computation costs and low efficiency. In this paper, we present a novel single- stage model via the Polar Coordinate Representation (PCR). In PCR, one human pose is integrated by a root joint, the angle of joint displacement, and the length of joint displacement. To locate the position of a person instance, previous one-stage methods required the creation of an additional root joint confidence map. To further reduce the computation cost of the one-stage method, we directly exploit the output map of the angle in the polar coordinate to replace the root joint confidence map. In our model, the redundancy prediction head of this root joint confidence map is dropped, which decreases the parameters of the network. For inference, our proposed network only needs one simple decoupling process to generate final poses. Our method achieves comparable performance to previous single-stage bottom-up methods, with lower computation cost and higher speed.