Inertia-1: An Open Exploration to a Unified Motion Foundation Model
15 hours ago
- Inertia-1 introduces a unified motion foundation model that overcomes dataset inconsistencies like varying sampling rates and sensor placements, enabling a single representation to adapt across different setups without retraining.
- The model leverages self-supervised learning on over 18 million hours of accelerometry data and maintains performance across body placements and sensor types (e.g., gyroscope, magnetometer), with multi-stream inputs improving accuracy and clustering.
- Key practical insights include optimal 30–60 second windows for tasks, the superiority of triaxial over vector-magnitude inputs, time-domain modeling for preserving gait cues, and robust activity recognition at low sampling rates like 1 Hz.