Multiple-branches faster rcnn for human parts detection and pose estimation

Abstract

In this work, we primarily address multiple people pose estimation challenge by exploring the performance of Faster RCNN on human parts detection. We develop a multiple-branches Faster RCNN model for our specific task of detecting persons and their parts. Our model can improve the performance of detecting human parts and the whole persons, meanwhile speeding up detection process with shared weights. A part-based method is proposed to estimate multiple people poses, bringing recent advances on object detection to this task. Experiments demonstrate that our model achieves better performance than the original Faster RCNN model on our task. Compared with other pose estimation approaches, our approach achieves fair or better results.