What Physics do Data-Driven MoCap-to-Radar Models Learn?

arXiv cs.LG / 5/4/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper investigates whether data-driven MoCap-to-radar models learn real underlying physics or merely produce plausible micro-Doppler spectrograms.
It introduces two physics-based interpretability metrics: one checks alignment with physics-derived Doppler frequency, and the other verifies preservation of the velocity–frequency relationship under velocity intervention.
The proposed metrics can be computed using only MoCap inputs and the model’s predictions, without needing measured radar data.
Experiments across multiple model architectures show that low reconstruction error does not necessarily imply physical consistency, with some models failing the physics-based metrics despite good error scores.
The analysis finds that temporal attention is crucial for transformer-based models to learn the underlying physics.

Abstract

Data-driven MoCap-to-radar models generate plausible micro-Doppler spectrograms, but do they actually learn the underlying physics? We introduce a physics-based interpretability framework to answer this question via two proposed complementary metrics: one measures alignment between model predictions and the physics-derived Doppler frequency, while the other tests whether predictions preserve the velocity-frequency relationship under velocity intervention. Both metrics require only MoCap input and model predictions, without access to measured radar data. Experiments across several model architectures reveal that low reconstruction error does not guarantee physical consistency: some, but not all, models achieve low error yet perform poorly on the two physics-based metrics. Further analysis shows that temporal attention is critical for transformer-based models to learn the underlying physics.