CCTV shoplifting detection: theft and normal shopping

This clip is a split-screen comparison from a high-angle camera in an electronics store. Display tables hold rows of phones. The left panel is titled Shoplifting and the right panel is titled Normal Shopping, and both show the same kind of scene so that the only difference is what the shopper does.

The subject, in a plaid shirt, picks up a boxed phone and examines it. The two panels then split. On the left the subject pushes the box inside the front of the jacket. On the right the subject puts the box back on the table and walks on. The two actions begin from almost the same pose, which is exactly why shoplifting is hard to detect from a single frame.

Each panel shows what the dataset labels contain. A line of caption text at the top gives the per-frame description written by a vision-language model. Blue boxes with confidence scores follow the people, and pose keypoints are drawn on their bodies. The caption, box and pose labels together let a model learn the sequence of the action and not only the final frame.

The dataset has more than 5,000 frames and over 100 video sequences, labelled with person boxes, 17-keypoint pose and VLM captions, and it is priced on request. A free sample on Kaggle includes 400 images and 8 clips.