Improvement in Multi-Person 2D Pose Estimation: Applying Polar Representation in OpenPose

Recent pose machines provide a relative accurate estimation in 2D real-time multi-person situations. In this work, we demonstrate an advanced open-pose design with a sequential stages of prediction and use of polar coordinate system. The main contribution of this paper is to denote a pose machine frame work based on the available open-pose model, which performs improvement in both efficiency and accuracy in image-dependent spatial models learning. We achieve this by considering additional information of image features with both a sequential structure of convolutional networks and the support of part affinity fields, as well as the advantages of using polar coordinate system, which efficiently predicting accurate estimates in multi-person cases. Our approach characterizes how the concept of part affinity fields can be used in key points connection. We perform competing methods on standard data sets including COCO data set, compare our result with several bottom-up approach and illustrate the result in straightforward ways.