Keypoint Detection for Identifying Body Joints using TensorFlow

Abstract

Keypoint Detection for the Body Joints is a process of finding and detecting all the important body joints of a human being. It takes an RGB image as input gives a list of keypoints as output. It can be applied for single-person body joints detection or multi-person body joints detection. The various architectures for detecting the human body joints can be compared using the Percentage of Detected Joints (PDJ), Object Keypoint Similarity (OKS) and mean Average Precision(mAP) evaluation metrics. The state-of-the-art body joint detection architectures like DeepPose, Higher Resolution Network and OpenPose architectures were studied and compared against each other. These architectures were evaluated on the COCO dataset for the 17 key body joints using the OKS and mAP evaluation metrics. The evaluation results on all the body joints were compared to each other for each architecture. The architectures were implemented using opensource implementations of the PyTorch and TensorFlow libraries. The HRNet architecture was the fastest and most accurate of all the architectures.