Downsizing Heatmap Resolution for real-time 3D Human Pose Estimation

In recent years, 3D human pose estimation based on deep learning is being developed and one of the important elements for these is realtime processing. However, the top-down method which is one of methods for 3D multi-person human pose estimation is not fast enough for real-time processing. In this paper, we propose to downsize the input resolution to the heatmap of RootNet and PoseNet, which are sub-frameworks of 3DMPPE framework. When heatmap resolution becomes smaller, computation cost and processing time decrease, and hence, we studied the trade-off relation between performance and processing time when heatmap resolution is reduced. When the input resolution was reduced from 256 to 128, the processing time in RootNet was decreased by 30%, while the performance was decreased by only 4%.