This is the official repository associated with the paper:
Online Fully Supervised Video Anomaly Detection
Authors Info:
Name: Vincenzo Carletti, Antonio Greco, Mattia Marseglia, Mario Vento
e-mails: vcarletti@unisa.it (V. Carletti); agreco@unisa.it (A. Greco); mmarseglia@unisa.it (M. Marseglia); mvento@unisa.it (M. Vento)
ORCID(s): 0000-0002-9130-5533 (V. Carletti); 0000-0002-5495-2432 (A. Greco); 0009-0009-0507-6884 (M. Marseglia); 0000-0002-2948-741X (M. Vento)
Submitted to Neurocomputing.
This repository contains the official implementation of the proposed online fully supervised video anomaly detection framework.
Unlike conventional fully supervised approaches that process complete videos offline, the proposed method operates in an online setting, where video streams are analyzed sequentially and alarm decisions are generated as new video chunks become available.
The framework combines:
- chunk-based video processing;
- frame-level anomaly score extraction;
- a continuity-aware mapping strategy that aggregates recent temporal evidence;
- real-time alarm generation with low latency.
The proposed formulation enables timely anomaly detection while maintaining the high recognition performance of fully supervised methods.
This repository will contain the resources required to reproduce the experiments presented in the paper, including:
- the code for training the proposed model;
- the code for performing online inference;
- the automatic temporal annotation tool used to generate event-level temporal annotations from surveillance videos.
Further documentation and usage instructions will be provided upon release.
- Online fully supervised video anomaly detection.
- Chunk-based video stream processing.
- Frame-level anomaly score prediction.
- Continuity-aware temporal mapping strategy.
- Low-latency alarm generation.
- Training and evaluation pipelines.
- Support for benchmark datasets.
- Reproducible experimental setup.
This repository also provides the temporal annotation resources introduced in the paper.
FS-UCF-Crime
FS-UCF-Crime is a temporally annotated extension of the widely used UCF-Crime benchmark. Each anomalous video is enriched with frame-level temporal boundaries, enabling the training and evaluation of fully supervised video anomaly detection methods.
Dataset link:
MIVIA-Crime
MIVIA-Crime is a new surveillance video benchmark with temporal anomaly annotations, specifically designed for cross-dataset evaluation. It enables the assessment of model generalization beyond the training distribution by providing a challenging test set acquired under different surveillance conditions.
Dataset link:
If you find this repository useful in your research, please consider citing our paper (citation will be added after publication).
For questions or further information, please contact the authors of the paper.