Visual dynamics models are often judged by how accurately they predict pixels, but that leaves a deeper question unanswered: does the predicted motion actually obey physical laws? The authors of a new arXiv preprint argue that geometric closure alone is insufficient, and that mechanical consistency must be measured separately.
To address this, they introduce a framework for measuring and diagnosing mechanical consistency in visual dynamics. The framework goes beyond traditional visual accuracy metrics, aiming to separate cases where motion looks right from cases where it is dynamically sound.
The approach includes an independent SI audit of rigid closed chains, which appears to be a key component in separating different aspects of consistency. The abstract does not specify what SI stands for, but the audit is presented as a way to diagnose mechanical consistency in a controlled setting.