eduzhai > Applied Sciences > Engineering >

Cascaded deep monocular 3D human pose estimation with evolutionary training data

  • Save

... pages left unread,continue reading

Document pages: 16 pages

Abstract: End-to-end deep representation learning has achieved remarkable accuracy formonocular 3D human pose estimation, yet these models may fail for unseen poseswith limited and fixed training data. This paper proposes a novel dataaugmentation method that: (1) is scalable for synthesizing massive amount oftraining data (over 8 million valid 3D human poses with corresponding 2Dprojections) for training 2D-to-3D networks, (2) can effectively reduce datasetbias. Our method evolves a limited dataset to synthesize unseen 3D humanskeletons based on a hierarchical human representation and heuristics inspiredby prior knowledge. Extensive experiments show that our approach not onlyachieves state-of-the-art accuracy on the largest public benchmark, but alsogeneralizes significantly better to unseen and rare poses. Code, pre-trainedmodels and tools are available at this HTTPS URL.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×