Balanced Feature Fusion for Grouped 3D Pose Estimation

3D human pose estimation by grouping human body joints according to anatomical relationship is currently a popular and effective method. For grouped pose estimation, fusing features of different groups together effectively is the key step to ensure the integrity of whole body pose prediction. However, the existing methods for feature fusion between groups require a large number of network parameters, and thus are often computational expensive. In this paper, we propose a simple yet efficient feature fusion method that can improve the accuracy of pose estimation while require fewer parameters and less calculations. Experiments have shown that our proposed network outperforms previous state-of-the-art results on Human3.6M dataset.