Grant: Discovery Grants Program - Individual

Graph representation learning (GRL), which focuses on developing techniques to learn expressive representations of graph-structured data, plays a crucial role in computer vision by enabling the modeling of relationships, capturing contextual information, combining visual and structural information, and facilitating generalization. GRL models have proven effective at encoding structural information, capturing complex spatial and temporal dependencies in graph-structured data. This makes them an ideal framework for addressing challenges in 3D human motion representation and analysis, given that the human skeleton can naturally be represented as a graph with body joints as nodes connected by edges. The practical applications of 3D human motion analysis span across a wide range of domains, including autonomous systems, virtual reality, healthcare, and sports. In healthcare, for instance, the analysis of 3D human motion holds significant importance in diagnosing movement disorders, monitoring patient progress, and improving rehabilitation outcomes. As a result, it contributes to enhancing the quality of care provided to patients. Through our prior research endeavors, we have developed GRL-based methods for diverse learning tasks, including 3D shape recognition, 3D human pose estimation, semi-supervised learning, and anomaly detection. Looking ahead, our focus will be on pursuing an integrated approach, aiming to develop a comprehensive framework for GRL that advances the state-of-the-art in 3D human motion representation and analysis tasks, including human motion prediction, human pose estimation, human mesh recovery, and other pertinent visual recognition tasks. The proposed research focuses on the following five interconnected research objectives: (1) explore and design GRL models tailored to the characteristics of 3D human motion data by leveraging graph neural networks, attention mechanisms, and other graph-based learning techniques to effectively capture the complex dependencies and interactions within the graph structure. This learning process will result in informative representations that encode both local and global contextual information; (2) investigate the efficiency and scalability of the developed algorithms with the aim of enabling their practical deployment in real-world scenarios; (3) develop methodologies that combine theoretical rigor with computational feasibility, enabling the analysis of 3D human motion and facilitating its practical applications; (4) explore the transferability of the developed GRL approaches to different visual data domains; and (5) exploit novel computer vision and geometry processing applications of the developed GRL solutions. Moreover, we anticipate that the outcomes of this research will have a broad impact on various downstream tasks, such as virtual try-on, opening new possibilities for realistic and immersive experiences in virtual environments.