Geometric Knowledge Distillation via Procrustes Analysis for Efficient Motion Sequence Classification

Motion sequence classification methods that rely on geometric approaches—such as Procrustes analysis and Dynamic Time Warping (DTW)—offer high accuracy but are often unsuitable for real-time applications due to their computational cost. In this paper, we present a novel geometric knowledge distillation framework that bridges the gap between accuracy and efficiency by transferring rich geometric insights from a Procrustes-DTW-based distance metric into a transformer-based neural network. By generating soft probability distributions from pre-computed Procrustes-DTW distances, we effectively guide the student model's training to preserve essential geometric properties like shape similarity, temporal alignment, and spatial transformation invariance. Our method enables fast and scalable motion sequence classification while retaining the benefits of geometric interpretability. We evaluate our framework on two benchmark tasks: sign language recognition using the SIGNUM dataset and human action recognition on UTD-MHAD. Results show that our distillation approach significantly improves classification accuracy over standard supervised learning and achieves dramatically lower inference time compared to traditional geometric methods—making it ideal for real-time motion understanding in wearable, robotic, and interactive systems.