QueryPose: SparseMulti-PersonPoseRegressionvia Spatial-AwarePart-LevelQuery

Thetwoindependent modelsleadtothenon-end-to-end pipeline, or called two-stage pipeline. Moreover, the human detector involves extra memory as well as computational cost. The bottom-up strategy [16, 17, 18, 19] uses the keypoint heatmap to locate all person keypoints at first and then assigns them to individuals via heuristic grouping process,asshowninFigure1(a).