A Plausibility Enhanced Data Augmentation Method for 3D Human Pose Estimation

In 3D human pose or shape estimation, the 2D-to-3D lifting method is usually applied directly or introduced as an assisted supervision strategy because its training dataset can be augmented easily from the 2D pose data, which thus leads to a more satisfactory result. However, sometimes the data may be generated with implausible poses, and so the training process is polluted accordingly. Related works focus on eliminating the obviously implausible poses but ignore the data being potentially implausible. This brings a risk to the training procedure that has a large number of iterations with data being generated continuously. In this paper, we propose a novel method to enhance the plausibility of the generated poses in detail. We first propose some limitations to distinguish the poses with less plausibility from others, and then further filter the implausible data by calculating the probability a pose being implausible through some factors of human poses. These means weaken the influence of implausible poses on the training effect without sacrificing the richness of training data. Experiments illustrate that the proposed method improves the performance of the trained model well.