Active Learning for 3D Human Pose Estimation: Combining the Strategies of Uncertainty and Diversity

3D human pose estimation is widely used in several applications such as VR/AR interactive games and sports analysis. However, the data annotation for training a 3D human pose estimation model is costly and time-consuming. Therefore, designing a practical system to label the 3D pose annotation efficiently is essential. In this work, we propose a two-phase active learning framework named LL-Core-Pose3D for 3D human pose estimation. The two-phase active learning framework combines uncertainty and diversity strategy to simultaneously query the information-rich and representative data. For the diversity-based strategy, we apply the Core-Set algorithm to calculate the relevant score, which is the similarity between unlabeled data points and labeled data points. For the uncertainty-based strategy, we adopt the idea of Learning Loss and design our own loss prediction model to predict the uncertainty score for 3D human pose estimation. In addition, We explore combining different features extracted from 2D and 3D as the input of human pose estimation models with our active learning framework. Our proposed strategy outperforms other existing active learning approaches with a smaller training dataset. Our technique can achieve the same performance as using the entire data with just using 66.98 % of the training dataset, which significantly saves the cost of training 3D human pose estimation models.