AVSS 2026, 22nd International Conference on Advanced Visual and Signal-Based Systems, 31 August-3 September 2026, Lecce, Italy
Violence recognition is dominated by abrupt temporal dynamics rather than static appearance. This creates a dilemma for surveillance systems: while RGB transformers
achieve strong performance through explicit temporal modeling, event cameras are increasingly considered as an alternative sensing modality better aligned with the motion-centric nature of violent actions. To investigate whether event-based sensing
can approximate powerful RGB pipelines, we first establish a strong RGB baseline by enhancing a frozen VideoMAE encoder with lightweight temporal modeling. We then compare this baseline against a state-of-the-art event-based model and the most
recent vision model on the UCF-Crime benchmark dataset using simulated RGB–event pairs. Our results show that event-based recognition remains competitive while being substantially more efficient. We further analyze the relationship between modalities
by fusing their prediction scores and show that RGB and event models capture complementary information. Finally, by unifying RGB and event data in a shared optical-flow domain, we find comparable performance when reduced to motion-only cues.
Overall, our findings indicate that while event cameras do not yet surpass strong RGB transformer pipelines, they constitute an efficient and complementary modality for violence understanding.
Type:
Conference
City:
Lecce
Date:
2026-08-31
Department:
Data Science
Eurecom Ref:
8859
Copyright:
© 2026 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.
See also:
