AdaptivePose: Human Parts as Adaptive Points

Multi-person pose estimation methods generally follow top-down and bottom-up paradigms, both of which can be considered as two-stage approaches thus leading to the high computation cost low efficiency. Towards a compact efficient pipeline for multi-person task, in this paper, we propose represent human parts points present novel body representation, leverages an adaptive point set including center seven human-part related instance more fine-grained manner. The representation is capable capturing various deformation adaptively factorizes long-range center-to-joint displacement delivers single-stage differentiable network precisely regress pose, termed AdaptivePose. For inference, our proposed eliminates grouping well refinements only needs single-step disentangling process form pose. Without any bells whistles, achieve best speed-accuracy trade-offs 67.4% AP / 29.4 fps with DLA-34 71.3% 9.1 HRNet-W48 on COCO test-dev dataset.