We propose the first real-time approach for the egocentricestimation of 3D human body pose in a wide range of unconstrained everydayactivities. This setting has a unique set of challenges, such as mobilityof the hardware setup, and robustness to long capture sessions withfast recovery from tracking failures. We tackle these challenges based on anovel lightweight setup that converts a standard baseball cap to a devicefor high-quality pose estimation based on a single cap-mounted fisheyecamera. From the captured egocentric live stream, our CNN based 3Dpose estimation approach runs at 60 Hz on a consumer-level GPU. Inaddition to the novel hardware setup, our other main contributions are:1) a large ground truth training corpus of top-down fisheye images and 2)a novel disentangled 3D pose estimation approach that takes the uniqueproperties of the egocentric viewpoint into account. As shown by ourevaluation, we achieve lower 3D joint error as well as better 2D overlaythan the existing baselines.
Download Everyday Baseball VR Rarl
Download File: https://demppomednu.blogspot.com/?iy=2vEz1F
2ff7e9595c
Comments