An In-Depth Analysis of 2D and 3D Pose Estimation Techniques in Deep Learning: Methodologies and Advances
Ruiyang Sun, Zixiang Lin, Song Leng, Aili Wang, Lanfei Zhao
Abstract
Pose estimation (PE) is a cutting-edge technology in computer vision, essential for AI-driven sport analysis, advancing technological applications, enhancing security, and improving the quality of life. Deep learning has markedly advanced accuracy and efficiency in the field while propelling algorithmic frameworks and model architectures to greater complexity, yet rendering their underlying interrelations increasingly opaque. This review examines deep learning-based PE techniques, classifying them from two perspectives: two-dimensional (2D) and three-dimensional (3D), based on methodological principles and output formats. Within each category, advanced techniques for single-person, multi-person, and video-based PE are explored according to their applicable conditions, highlighting key differences and intrinsic connections while comparing performance metrics. We also analyze datasets across 2D, 3D, and video domains, with comparisons presented in tables. The practical applications of PE in daily life are also summarized alongside an exploration of the challenges facing the field and the proposal of innovative, forward-looking research directions. This review aims to be a valuable resource for researchers advancing deep learning-driven PE.