Geometric Feature Relationship-Based Knowledge Distillation for Ground Reaction Force Estimation.

Wearable sensor-based gait analysis has broad applicability across healthcare, rehabilitation, and injury prevention. Ground reaction force (GRF) provides key insight into human-ground interactions but is typically measured with instrumented treadmills and force plates that are costly and confined to laboratory settings. Insole sensors offer a portable alternative, but their sensings are often noisy and susceptible to environmental interference. Deep learning can mitigate these issues, but high accuracy usually demands substantial compute, hindering real-time and on-device utilization. With these, knowledge distillation can be adopted as a promising solution to generate a compact model. However, many feature- and relation-based distillation methods compute teacher-student relations using fixed similarity or metric operators, such as inner products and Euclidean distances. directly on intermediate representations. Although these have been widely used in feature matching, they do not explicitly learn a shared nonlinear feature geometry in which teacher and student relations are compared. This can also be limiting when the teacher and student have different capacities, channel widths, and representation distributions. To address these issues, we propose Geometric Feature Relationship-Based Knowledge Distillation for Ground Reaction Force Estimation, GFKD. During distillation, intermediate features from the teacher and student are leveraged in a teacher-student shared assistant module (TSM) that operates a shared intermediate alignment between their representations. TSM computes nonlinear geometric relations and encourages the student to capture them, providing geometry-aware supervision instead of general direct feature matching. In our experiments, students distilled by GFKD outperforms recent baselines for GRF estimation, showing the robustness of the proposed method in both accuracy as well as resource-efficiency.