Environment Adaptive 3D Pose Estimation Model and Learning Strategy

Recently, 3D pose estimation models using deep learning structures have begun to show outstanding performance. However, the performance is guaranteed only for the general pose included in public databases. In other words, most estimation models sometimes show degraded results when a given video contains uncommon poses from specific situations such as exercise and dance. This problem arises from the limitation of the pose diversity of public databases. We propose a novel estimation model calibration (EMC) framework for environment adaptive 3D pose estimation to solve this problem. This framework aims to calibrate well-trained existing pose estimation models from public databases to suit the environment. To achieve this goal, the framework uses target data to analyze the problems of existing estimation models. Subsequently, the proposed ergonomic model handler generates a calibration dataset by directly correcting the problem caused by the target data. Using the generated calibration dataset, we calibrate the existing pose estimation model. In this paper, we provide various experimental results of pose estimation for verification of the proposed framework. Experimental results demonstrate performance improvements qualitatively and quantitatively in specific poses and show the efficiency of estimation model calibration.