UC Berkeley And Meta AI Researchers Propose A Lagrangian Action Recognition Model By Fusing 3D Pose And Contextualized Appearance Over Tracklets