In the field of human pose estimation, regression-based methods have been dominated in terms of speed, while heatmap-based methods are far ahead in terms of performance. How to take advantage of both schemes remains a challenging problem. In this paper, we propose a novel human pose estimation framework termed DistilPose, which bridges the gaps between heatmap-based and regression-based methods. Specifically, DistilPose maximizes the transfer of knowledge from the teacher model (heatmap-based) to the student model (regression-based) through Token-distilling Encoder (TDE) and Simulated Heatmaps. TDE aligns the feature spaces of heatmap-based and regression-based models by introducing tokenization, while Simulated Heatmaps transfer explicit guidance (distribution and confidence) from teacher heatmaps into student models. Extensive experiments show that the proposed DistilPose can significantly improve the performance of the regression-based models while maintaining efficiency. Specifically, on the MSCOCO validation dataset, DistilPose-S obtains 71.6% mAP with 5.36M parameters, 2.38 GFLOPs, and 40.2 FPS, which saves 12.95×, 7.16× computational cost and is 4.9× faster than its teacher model with only 0.9 points performance drop. Furthermore, DistilPose-L obtains 74.4% mAP on MSCOCO validation dataset, achieving a new state-of-the-art among predominant regression-based models. Code will be available at https://github.com/yshMars/DistilPose.

Keypoint heatmaps, which produce peak values of Gaussian distributions for individual human joints, are crucial components in 2D human pose estimation. However, existing merits using keypoint heatmaps are usually defeated in body occlusion, resulting in inaccurate joint predictions. We consider that the failure is mainly due to keypoint heatmaps’ insufficiency for distinguishing the joints from two occluded bodies. Therefore, in this paper, we propose a method termed SkeletonMap (SMap), which introduces the prior knowledge of body structure to constrain relative connection of joints. As an extension of keypoint heatmaps, SMap can be efficiently plugged into existing 2D human pose estimation models with negligible increase in computational cost. Extensive experiments are conducted to show the effectiveness and generalization of SMap. Without bells and whistles, SMap brings a significant performance boost to the existing heatmap-based 2D human pose estimation models. On the MPII dataset, SMap improves SimpleBaseline (ResNet-152) from89.7 [email protected] to 90.4, and HRNet (W32) from 90.5 to 90.9. On the COCO dataset, SMap improves SimpleBaseline (ResNet-152) from 72.4 AP to 73.9. On the more challenging OCHuman dataset, SMap improves HRNet (W32) from 61.9 AP to 64.5, achieving 2.6 AP gains. We hope our simple and efficient approach will serve as a solid component for future research in 2D human pose estimation.