Human Pose Estimation with Combined Feature Maps and Joint Embeddings
Tianyuan Han, Ganyu Huang, Chunhui Li, Liping Shen · 2023
In recent years, Human Pose Estimation task, a typical computer vision task, still remains room for further improvement in overcoming complex scenarios (e.g., occlusion), data flaws (e.g., motion blur), etc. Recent researches often equip deep learning models with human prior knowledge to obtain enhanced accuracy. In this study, we present a novel model that simultaneously maintains both joint graph embeddings and feature maps as two separated branches and communicates information between them, aiming to enable the model to leverage both long-range semantic relationships and local textile information effectively. We evaluate our experiment on PoseTrack18 dataset and compare its performance with existing state-of-the-art methods. Our model achieves a mean Average Precision (mAP) of 80.2, which, while not out-performed, is competitive with existing state-of-the-art methods. As most previous methods solely use either feature maps or joint embeddings to generate their prediction, further research will be conducted to explore if parallelly combining feature maps and joint embeddings with them can further enhance the result.