Validate the capacity–data-size interaction of the event-boundary gate

Determine whether the event-boundary gate's effect on video anomaly detection accuracy is governed by an interaction between model capacity and the amount of training data, by evaluating the gated and ungated causal state-space detector on a substantially larger third dataset such as ShanghaiTech.

Background

The causal diagonal state-space detector uses an input- and state-dependent event-boundary gate to modulate its effective decay and react rapidly to changes in the input. Ablation results differ across the two evaluated datasets: disabling the gate improves accuracy on the smaller UCSD Ped2 training set but substantially reduces accuracy on the larger CUHK Avenue training set. The paper interprets this reversal as possible evidence that the gate's usefulness depends on the interaction between model capacity and available training data, but emphasizes that the conclusion is based on only two datasets.

The authors state that a substantially larger third training set is required to test this hypothesis properly. ShanghaiTech was intended for that evaluation, but its raw data had not been fully obtained, leaving the comparison unresolved.