Generalization of the appearance-to-damping mapping beyond controlled impact footage

Determine whether the trained projection head learns a general appearance-to-damping mapping or associations tied specifically to controlled drumstick footage, extending evaluation beyond the Greatest Hits dataset and its single visual domain.

Background

The study evaluates acoustic-to-visual transfer using the Greatest Hits dataset, which contains controlled videos of objects struck with a drumstick and covers 17 material categories. The visual student is a frozen CLIP encoder augmented with a trainable projection head, while the audio teacher provides the distillation target.

The reported results therefore establish transfer only within this dataset and visual setting. It remains unresolved whether the projection head has learned a broadly applicable mapping from visual appearance and dynamics to damping, or whether it has exploited associations specific to the recording setup and controlled impact footage.