3D Pose Estimation based on Deep Learning without Real World Data Training
Chen-Ruei Sin, Wei-Liang Lin · 2020
This paper proposes a new methodology to carry out the 3D pose estimation of an object. We assume a single lens camera environment, and use a game engine to generate virtual world datasets in order to assist a deep learning model. Using only the virtual world datasets to conduct training has been tested with the real world data and the accuracy can be achieved up to 84.22% with 30M parameters.