What is a video dataset for computer vision?

A video dataset for computer vision is a collection of short clips paired with per-frame or per-clip labels — bounding boxes, keypoints, or action classes — used to train models that reason about motion over time rather than single frames. Typical use cases are human activity recognition, action recognition, fall detection, and CCTV anomaly detection.

How is a synthetic video dataset annotated?

Synthetic video is annotated at render time. Because the pipeline knows every character's skeleton, every object's bounding volume, and every action label, it can output pixel-accurate bounding boxes, 17-keypoint skeletons, and action-class labels for every frame automatically. There is no manual labeling step and no inter-annotator disagreement.

What is the difference between human activity recognition and action recognition?

Action recognition usually refers to short, atomic movements — a wave, a jump, a punch — classified from a few seconds of video. Human activity recognition (HAR) is broader and includes longer, compound behaviors such as loitering, shoplifting, or a fall event that unfolds over multiple seconds. HAR datasets are typically longer clips with temporal boundaries.

Are synthetic video datasets GDPR compliant?

Yes. Every person in a synthetic video is computer-generated, so no real biometric data is captured, stored, or processed. That removes the consent, retention, and lawful-basis burdens that come with training on real CCTV footage under GDPR or the EU AI Act's rules on biometric training data.

How do I get a custom video dataset for my use case?

Simuletic generates custom synthetic video datasets on request. Send the behavior, camera type, environment, and label format you need and we produce clips with per-frame annotations, typically in days rather than months.

Video Datasets for Computer Vision

Why video is the harder problem

Image datasets have had a decade of open-source generosity. Video datasets have not. A single labeled fall event, weapon draw, or shoplifting sequence can take a labeler thirty minutes to annotate frame by frame. Real footage is also legally and ethically radioactive: care-home CCTV, transit-station platform cameras, and store surveillance are covered by privacy law almost everywhere they matter.

So the teams shipping serious video models — safety, retail, eldercare, transit, defense — are quietly moving to synthetic video. The clip below is one of ours: a synthetic person collapsing from cardiac arrest. No actor was hurt, no CCTV was scraped, and every frame comes with a labeled skeleton and action class.

Synthetic 6-second clip: person collapsing from cardiac arrest. Rendered, not filmed — the exact edge case fall-detection and CCTV anomaly models need.

What counts as a video dataset for CV

Any collection of clips with temporal labels: per-frame bounding boxes, per-frame skeletons, or clip-level action classes. The three families you'll see in production are:

Action recognition — short atomic clips labeled with one class (wave, kick, fall).
Human activity recognition (HAR) — longer sequences with temporal boundaries around compound behaviors like loitering or a fall event.
Video anomaly detection — normal-versus-abnormal training where the "abnormal" class covers everything from a fall to a fight to an abandoned bag.

The categories that matter in 2026

These are the video-dataset categories with active buyer demand — safety, security, retail loss, transit, and defense. Each links to a live Simuletic dataset or write-up.

Fall detection video

Person collapsing, slipping, or lying down. Used in eldercare, hospitals, and workplace safety.

Read the fall detection case study

Aggression and violence detection

Fights, weapon draws, aggressive postures classified from short clips.

See the aggression detection dataset

Shoplifting and retail anomaly

Concealment, sweep-and-grab, and loitering behaviors captured from overhead CCTV.

Read the shoplifting dataset write-up

ATM robbery and hold-up

Weapon presentation, forced withdrawals, and getaway behavior at ATM cameras.

See the ATM robbery dataset

Smoke and fire early warning

Early smoke plumes and flame onset across indoor CCTV and outdoor wildfire perspectives.

Read the smoke and fire post

Pedestrian and behavior attributes

Age, clothing, carried items, and gait for downstream re-identification and search.

See the pedestrian attribute dataset

Why synthetic video beats scraped video

Rare events on demand. A cardiac collapse, a weapon draw, a fire igniting. Waiting for these to happen on real CCTV is not a data strategy.
Perfect temporal labels. Every frame's skeleton, bounding box, and action class comes out of the renderer, not a labeling contract.
No PII. Nothing to redact, no consent to chase, no retention clock to run.
Camera match. You can render at your exact deployment resolution, mount height, and lens — not adapt to whatever the open dataset happened to use.

How Simuletic ships video data

Every synthetic video dataset we release comes with per-frame YOLO bounding boxes, optional 17-keypoint COCO skeletons, and clip-level action labels. The current library covers falls, aggression, shoplifting, ATM robbery, smoke and fire, drone threats, and pedestrian attributes — plus custom scenarios on request. Full lists on the datasets page.

Need a video dataset that doesn't exist yet?

Send us the behavior, the camera, and the label format. We generate synthetic clips with per-frame annotations, usually in days.

Request a custom video dataset Browse all datasets CCTV video anomaly guide