HumanTracker: A New Benchmark for Human-Aligned Motion Tracking Evaluation
A new benchmark called HumanTracker has been launched by researchers to enhance the evaluation of humanoid motion tracking, ensuring it is both perceptually aligned and scalable. This benchmark comprises around 153 hours of optical motion data from various professional performers, categorized into four motion families with detailed text labels for precise analysis. Additionally, the team introduces HumanScore, a metric based on preferences, trained on 12,000 motion pairs featuring 24,000 distinct motions. HumanTracker aims to overcome the shortcomings of existing evaluation techniques, which often conflict with human perception by concentrating on kinematic errors while neglecting physical artifacts like unstable support and improper contacts. This work, emphasizing contact-rich and long-horizon behaviors, is documented in a paper on arXiv under identifier 2608.13555.
Key facts
- HumanTracker benchmark contains approximately 153 hours of optical motion trajectories.
- The benchmark is organized into four motion families with text labels.
- HumanScore is a preference-aligned metric trained on 12K motion pairs containing 24K motions.
- The benchmark addresses limitations in current evaluation methods that disagree with human perception.
- Current methods often miss physical artifacts like unstable support and incorrect contacts.
- HumanTracker aims to stress contact-rich, long-horizon behaviors.
- The paper is available on arXiv with identifier 2608.13555.
- The benchmark is designed to be scalable and perceptually aligned.
Entities
Institutions
- arXiv