A Multi-view 3D Human Pose Estimation Algorithm Based On Positional Attention

Abstract

With the development of CNNs, the human pose estimation research has made great progress, but there is still a problem: the relationships of the human each joint location are not well exploited in previous CNNs-based methods. Considering the order of global spatial information and human body location information, we propose a multi-view 3D human pose estimation algorithm based on position attention. In 2D detection stage, position coding is adopted to rebuild the image in the global space position relation. The attention mechanism can model the relationship between various channels and capture feature maps the dependencies between the horizontal and vertical direction, and the details are mined from the feature location relationship to generate high-quality feature maps. In the last stage of feature extraction, adjacent view features are used to enhance the spatial expression ability of feature images, so as to better solve occlusion and oblique view. Experiments on the Human3.6M data set show that when using Resnet-50 as the backbone network and 256×256 of the image size, the average joint error of our algorithm is reduced to 25.2mm, which reaching the competitive result.