Human Pose Estimation: Benchmarking Deep Learning-based Methods

Human posture estimation (HPE) aims to detect human body parts and generate a demonstration of the human body with incoming data such as images and video. HPE has lately gained much attention as shown numerous applications in the real world. State-of-the-art human posture estimation methods may benefit from deep learning. In this view, this study aims to examine and assess deep learning-based human posture estimation algorithms. To describe and analyze recent research, multiple deep learning-based human posture estimation approaches (OpenPose, ViTPose-B, HRNet, AlphaPose, DenseNet, EfficientPose, DensePose, Hourglass,
$4^{*}\text{RSN}-50$
) deployed on the COCO and MPII datasets. This paper examined human posture using evaluation measures like average precision (AP) and probability of correct key points (PCK). Therefore, ViTPose-B methods outperform AP on the COCO dataset & Hourglass achieves better PCK on MPII. The comparative analysis helps to understand the applicability of different techniques for human pose estimation. Furthermore, unresolved issues and future research challenges of the human pose are mentioned.