An Effective 3D Human Pose Estimation Method Based on Dilated Convolutions for Videos

Abstract

3D human pose estimation is a challenging problem due to the diversity of poses, human appearance, clothing, occlusion, etc. In this work, we split the problem into two stages, 2D human pose estimation and 3D pose recovery, and address it by a network based on dilated convolution for videos. We introduce pruning layer to prevent overfitting, which performs better than dropout, because the strategy of pruning is not to drop nodes randomly, but to choose the lower-weight ones. We also employ quantization to accelerate it in the smart shop environment, which gains a trade off between performance and computational complexity, and the accuracy loss is in an acceptable range. In addition, labeled human pose datasets are so limited and expensive especially for 3D poses, we propose a 3D human pose dataset named HRI-I in a smart shop environment, which contains more than 16k poses, 26 people and 6 scenarios of walking, hunkering, fetching objects, etc. We train and test our model on the HumanEva-I, Human3.6M and our proposed HRI-I, it demonstrates that the proposed method is efficient and effective.