Unsupervised 3D Human Pose Estimation

Aug 30, 2020
PDF
Model Predictions

  • Wrote data-loaders and modeled the architecture for kinematic-structure preserving, unsupervised 3D pose estimation framework to effectively disentangle pose, foreground and background appearance information.

  • Reduced MPJPE by as high as 40% (semi-supervised) and 15% (unsupervised) on datasets like Human3.6M, 3DHP, LSP and 3DPW.

  • Advisors: Venkatesh Babu and Jogendra Nath Kundu

Deep Learning Computer Vision
Benedict Florance Arockiaraj
Benedict Florance Arockiaraj
ML Engineer

My research interests are at the juncture of deep learning and computer vision.