CCTV Shoplifting Detection Dataset (YOLO + VLM)

Retail shrink is a $100B+ problem and most off-the-shelf vision models can't tell a shopper from a shoplifter — because both look identical until the moment of concealment. This dataset is purpose-built for that gap. Scenes are rendered from realistic high-angle in-store CCTV perspectives across electronics stores, grocery aisles, apparel sections, and convenience layouts, with person bounding boxes plus 17-keypoint pose annotations capturing the subtle body language of pocketing, bag-stashing, and shelf-sweep behaviors. The package ships in two formats: YOLO labels for spatial detection and VLM-style natural language captions for fine-tuning vision-language models on behavioral context. Video sequences are included so temporal models (3D CNNs, video transformers) can learn the time signature of a concealment event — the pause, the glance, the hand returning empty. 100% synthetic — no real customers, no biometrics, no GDPR exposure. The open-source sample on Kaggle includes 400 frames and 8 video clips; the full package contains 5,000+ frames and 100+ video sequences covering diverse store types, lighting conditions, and shoplifting techniques.