Transfer Learning for Human Motion Prediction: Improving Accuracy Under Data Scarcity

Abstract

Human motion prediction (HMP) consists of anticipating future human poses based on motion history. While recent approaches achieved strong performance on benchmark datasets, their applicability to real-world scenarios remains limited due to the scarcity of context-specific motion data. In this work, we propose a unified, model-agnostic transfer learning (TL) framework for HMP to improve prediction accuracy under data-scarce target domains. The framework leverages pretraining on a large dataset of generic motions to learn a transferable motion prior, followed by systematic adaptation to context-specific actions using different TL techniques. We present a comprehensive evaluation protocol to characterize prediction accuracy across multiple representative HMP architectures, target datasets, and TL strategies and to evaluate training efficiency by introducing dedicated TL metrics. Beyond performance gains, we analyze the impact of source-target dataset distance and model complexity on TL effectiveness. The results show that TL consistently improves accuracy across all models and target datasets, reducing error up to 32.95% compared with training from scratch. In addition, TL enhances convergence speed, significantly reducing training epochs. These findings demonstrate TL as an effective and versatile approach for improving HMP under data scarcity, and provide methodological insights for its deployment in real-world context-specific applications.