Interaction Representations for Early Violence Detection

Work on interaction-aware video models generally reports that adding an interaction module improves performance. It rarely separates that improvement from the backbone, the training schedule, or the evaluation protocol — so the size of the contribution attributable to the interaction representation itself is usually unknown.

This is framed as a measurement study, not a detector. Early detection forces a model to commit before an event has finished, on partial evidence, which makes it a sharp setting in which to ask what a representation is actually contributing. The value of the answer is methodological: it tells you whether the module is doing work, or whether the backbone was.

Hold the backbone, the data and the evaluation protocol fixed, and vary only the interaction representation. The reported quantity is the attributable contribution under those fixed conditions — a benchmark-scoped number, not a general capability claim.

This is empirical research on representation learning conducted on benchmark datasets. It is not a deployable surveillance system and is not presented as one. Systems that classify human behaviour from video carry well-documented risks — distributional bias, error costs that fall unevenly on the people being observed, and function creep once deployed. Any claim from this work is scoped to the benchmarks it was measured on.

Characterising attributable contribution as a function of how much of an event has been observed, and identifying the point at which interaction representations stop helping.