Human pose estimation using artificial intelligence

Abstract

Human Pose Estimation (HPE), is recently a popular research topic in computer vision field. The object of this task is to generate a skeleton-like representation of a human body and then process it further for task-specific applications. Due to the outstanding performance of vision transformer in other Computer Vision (CV) tasks, its application to HPE problems and ability to surpass Convolutional Neural Network (CNN) models is gaining attention. Apart from the deep learning model, data processing in HPE cannot be ignored. Through the coordinate system transformation in common HPE solutions, the accuracy degrades when flipping strategy is applied in inference. The bias is derived from using resolution measured in pixels instead of size measured in unit length when performing resizing transformation. In this dissertation, TransPose, one state-of-the-art model introducing transformer for HPE is studied. The model is trained and tested on COCO benchmark. The modified TransPose model with unbiased approach is experimented as well. Based on the results, the dependency function of attention map is analyzed. The insights gained from this work provides valuable insight for future exploration of transformer applications and data processing in human pose estimation.