Shoplifting Video Dataset
This synthetic video dataset offers simulated shoplifting and normal behavior scenes captured in a real environment. Ideal for detecting human actions through deep learning.
182 files (.mp4 videos) — 640×480 resolution — 30 FPS — 2 classes
CC BY 4.0
Description
The dataset Shoplifting Video contains video clips captured in a computer vision laboratory, representing two behaviors: “normal” (walking, product inspection) and “shoplifting” (hiding objects in bags or under clothing). The videos are recorded with a 32 MP camera, in 640×480, at 30 frames/s.
What is this dataset for?
- Train models to detect suspicious behavior in stores
- Testing action recognition algorithms on video streams
- Create prototypes of intelligent surveillance systems
Can it be enriched or improved?
Yes, we can complete this dataset with other camera angles or diversify the scenarios (types of theft, interactions between people). Additional annotations (bounding boxes, keypoints) could also be added for multi-task models (action + location).
🔎 In summary
🧠 Recommended for
- Computer vision researchers
- Video surveillance projects
- Anomaly detection
🔧 Compatible tools
- OpenCV
- PyTorch Video
- Detectron2
- Kinetics-based pipelines
💡 Tip
For more robustness, couple this dataset with real flows or flows generated synthetically via GaNS or 3D simulations.
Frequently Asked Questions
Does this dataset contain frame-by-frame annotations?
No, the videos are simply labeled as “normal” or “stolen.” There are no temporal or spatial annotations.
Is it possible to use this dataset for real-time detection?
Yes, it can be used as initial training for surveillance camera action detection models.
Is this dataset suitable for supervised learning?
Yes, videos are classified by binary label, which makes it possible to train a supervised video classifier.
