GAT-POSE: Graph Autoencoder-Transformer Fusion for Future Pose Prediction

Armin Danesh Pazho, Gabriel Maldonado, Hamed Tabkhi

Abstract: Human pose prediction, interchangeably known as human pose forecasting, is a daunting endeavor within computer vision. Owing to its pivotal role in many advanced applications and research avenues like smart surveillance, autonomous vehicles, and healthcare, human pose prediction models must exhibit high precision and efficacy to curb error dissemination, especially in real-world settings. In this paper, we unveil GAT-POSE, an innovative fusion framework marrying the strengths of graph autoencoders and transformers crafted for deterministic future pose prediction. Our methodology encapsulates a singular compression and tokenization of pose sequences through graph autoencoders. By harnessing a transformer architecture for pose prediction and capitalizing on the tokenized pose sequences, we construct a new paradigm for precise pose prediction. The robustness of GAT-POSE is ascertained through its deployment in three diverse training and testing ecosystems, coupled with the utilization of multiple datasets for a thorough appraisal. The stringency of our experimental setup underscores that GAT-POSE outperforms contemporary methodologies in human pose prediction, bearing significant promise to influence a variety of real-world applications favorably and lay a robust foundation for subsequent explorations in computer vision research.