A lightweight 2D Pose Machine with attention enhancement

Abstract

Pose estimation is a challenging task in computer vision that has many applications, as for example: in motion capture, in medical analysis, in human posture monitoring, and in robotics. In other words, it is a main tool to enable machines do understand human patterns in videos or images. Performing this task in real-time while maintaining accuracy and precision is critical for many of these applications. Several papers propose real time approaches considering deep neural networks for pose estimation. However, in most cases they fail when considering run-time performance or do not achieve the precision needed. In this work, we propose a new model for real-time pose estimation considering attention modules for convolutional neural networks (CNNs). We introduce a two-dimensional relative attention mechanism for feature extraction in pose machines leading to improvements in accuracy. We create a single shot architecture where both operations to infer key points and part affinity fields share the information. Also, for each stage, we use tensor decompositions to not only reduce dimensionality, but also to improve performance. This allows us to factorize each convolution and drastically reduce the number of parameters in our network. Our experiments show that, with this factorized approach, it is possible to achieve state-of-art performance in terms of run-time while we have a small reduction in accuracy.