Researchers have built a new video violence detection model that trades heavyweight Transformer compute for a more efficient state-space backbone.
FuseMamba-VD splits its work across two branches: one reads spatial features from individual frames while the other tracks how those features change over time. A gating mechanism continuously feeds spatial signals into the temporal branch, letting the model catch violent events even when camera angles are awkward or lighting is poor. The team also stitched together four existing datasets - RWF-2000, RLVS, SURV, and VioPeru - into a single benchmark with strict train/test separation to make comparisons harder to game. On that benchmark and on the separately introduced DVD dataset, the model claims state-of-the-art accuracy.
The efficiency angle is the part worth watching. Transformers scale badly on long video clips because attention cost grows with sequence length; state-space models sidestep that by processing sequences in linear time. If the efficiency claims hold under real deployment conditions, the architecture could run on the kind of edge hardware that actually lives inside a camera housing rather than in a data center.
Surveillance AI has a long record of performing well on academic benchmarks and poorly in the field - especially on underrepresented demographics - so the benchmark construction choices here matter as much as the accuracy numbers. The code and pre-trained weights are public on GitHub, which at least lets others check the work.