Abstract: Although significant progress has been achieved in monocular3D human pose estimation, the correlation between body parts andcross-view geometry consistency have not been well studied. In this work,to fully explore the priors on body structure and view-relationship for3D human pose estimation, we propose an anatomy and geometry constrainedone-stage framework. First of all, we define a kinematic structuremodel in deep learning framework which represents the joint positionsin a tree-structure model. Then we propose bone-length and bone-symmetrylosses based on the anatomy prior, to encode the body structureinformation. To further explore the cross-view geometry information,we introduce a novel training mechanism for multi-view consistencyconstraints, which effectively reduces unnatural and implausible estimationresults. The proposed approach achieves state-of-the-art results onboth Human3.6M and MPI-INF-3DHP data sets.