Overview
The SUT Anomaly Detection Dataset was collected in an indoor surveillance setting by Habib Ebadi Namin at Shahrood University of Technology to provide a high-resolution benchmark for single-scene video anomaly detection. The dataset comprises 86 videos with a total of 26,483 frames, split into 59 training videos (14,642 frames) and 27 testing videos (11,841 frames). All frames have a spatial resolution of 1920 × 1080 pixels.
The training set contains only normal activities and includes natural background variations, such as changes in indoor illumination and scene decoration, to improve model robustness. The testing set includes both normal and anomalous events with frame-level ground-truth annotations for quantitative evaluation. Anomalies in the dataset fall into two challenging categories: appearance-only anomalies, such as the displacement of large objects (e.g., tables and chairs), and short-term motion anomalies, such as running. Running events occur both toward/away from and across the camera, introducing significant object-scale variations due to perspective.
Dataset Statistics
| Training | Testing | Total | |
|---|---|---|---|
| Videos | 59 | 27 | 86 |
| Frames | 14,642 | 11,841 | 26,483 |
| Resolution | 1920 × 1080 | ||
| Color | RGB | ||
| Format | PNG / MP4 | ||
| Frame rate | 25 FPS | ||
| Annotation | — | Frame-level | — |
Training split contains only normal events; testing split includes frame-level anomaly labels.
Sample Frames
Citation & Acknowledgments
BibTeX template will populate here.
Special thanks to Dr. Alireza Ahmadifard for supervision, and to Dr. Reza Kharghanian for feedback and constructive advice.