Disagreement-aware late fusion of 3DCNN and MViT for clip-level video anomaly detection
No Thumbnail Available
Date
2026-06-30
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Springer Science and Business Media B.V.
Abstract
We propose Adaptive Late Router with Temperature Scaling (ALeRT), an inference-only, disagreement-aware late fusion framework for clip-level video anomaly detection. ALeRT combines two complementary experts: MMAnet, a lightweight multi-scale 3D CNN for fine-grained motion dynamics, and MViT, a Kinetics-pretrained multiscale Vision Transformer for long-range spatiotemporal modeling, without requiring joint retraining. Model outputs are calibrated using temperature scaling and fused via a Product-of-Experts (PoE) to reinforce agreement, while disagreement is resolved using confidence margins and per-class Positive Predictive Value (PPV). Temporal consistency is further enhanced through exponential moving average (EMA) smoothing. The framework is evaluated across multiple random seeds, reporting mean performance with low variance, demonstrating strong robustness and generalization. Experiments on LPS (Loss Preventive Surveillance), UCF-Crime-Clips (UCF-CC), and Virat2-RC (ReClassified) datasets show consistent improvements over individual backbones. Notably, ALeRT achieves these gains with only 3.1 ms fusion overhead per clip, enabling real-time inference (5.7–6.0 events/s, maximum of 300 FPS(FPS (frames per second) indicates the number of video frames processed per second.)) on commodity GPUs. These results highlight the effectiveness of disagreement-aware fusion for scalable and deployment-ready video anomaly detection.