Synthetic CCTV datasets for video analytics Rare events are hard to film.So we built the training data.
Your detector has seen millions of people walking. It has seen almost no one fall onto a railway track. We simulate those events, label every frame and deliver 37,440 sequences with 12 cameras on each event, a steady 30 frames per second and a commercial licence.
Free sample and datasheet for every dataset. Prices from 3,000 EUR net.
Built for teams that train their own detection models
We have produced datasets for
through an agency partner
Falls
952 events and 12 cameras per event, with fall phases on every frame.from 18,000 EUROrder from 18,000 EURFights
1,033 events, 10 kinds of fight and 10 look-alikes such as hugs and high fives.from 3,000 EUROrder from 3,000 EURRail
1,135 events across 3 station scenes, with falls and pushes onto the track.from 3,000 EUROrder from 3,000 EURFire and smoke
89 video clips, 10,822 annotated frames, with steam and fog labelled as look-alikes.from 2,900 EURSee the datasetBuilt for teams that ship detectors.
You do not need more footage of nothing happening. You need the rare event, labelled, in the numbers your model requires. These packs give you that.
Video analytics vendors
Your model needs the fall, the fight and the track intrusion, not another hour of people walking. Train on thousands of labelled events plus look-alikes that cut unwanted alarms. Test the labels on the free sample before you order.
Rail and metro safety teams
You cannot stage a push onto the track. We simulated 13,620 sequences from platform cameras, with a zone label on every frame: platform, beyond line, edge, track.
Robotics, research and safety startups
Train pose classifiers on 17-keypoint skeletons from 12 cameras per event. Splits follow the base motion, so your test set stays honest. Research editions for fights and rail start at 3,000 EUR.
Six things you can check before you buy.
We do not ask you to take our word. Each point is in a datasheet or in the free sample.
Labels on every frame
A box and a fall phase per person on each frame. Fights add roles and contact events, rail adds a zone.
12 angles on every event
The same event from 12 virtual cameras. Your model learns angles you never filmed.
A steady frame rate
Measured, not assumed. Fights ran at 26.8 fps in the first release. All three packs run at 30.
Look-alikes built in
Falls hold 50.5% non-fall motion, fights about 45% hard negatives such as hugs. Your detector learns to stay quiet.
Quality control we publish
Fight clips with a pillar overlap are gone. In rail, 40 of 40 stumble clips now end upright.
A licence you can read
Full text on every page. Standard and Enterprise allow commercial use. Training generative AI is banned.
What changed in the latest releases.
We fixed what our own quality control found. These numbers come from our logs and datasheets, and you can check them against the free samples.
Falls: genuine near misses
Sequences labelled near_miss, so your model learns a stumble from a fall.
Fights: people overlapping a pillar
Clips where a person overlaps a pillar for at least 1 second, or a body part sits inside one.
Rail: stumble and recover
Clips of edge_stumble_recover where the person returns upright.
Size of each pack
Labelled sequences in the current releases.
What you can check in each datasheet: per-frame labels, 12 virtual cameras per event, even 30 fps, automated motion quality control and the licence text. Previews are renders from the same simulations. The packs hold skeleton data and labels, not video.
Your model needs the event.
Real footage rarely has it.
Falls, fights and early smoke are rare on camera and slow to label. Synthetic data gives your team those events now, with labels included.
From this page to your training run.
Test the free sample.
Open a dataset, read the datasheet and try the sample on your own model.
Choose a licence.
Send us your order. We reply with an order form and invoice.
Start training.
We send a secure download link within 3 business days of payment.
Your scene. Your cameras. Your classes.
Our packs cover falls, fights and railway platforms. Your cameras see something else? Tell us. We build the scene, simulate the events and look-alikes you name, label every frame and deliver in your format.
- Your scene and camera layout
- The events you need, and the unwanted alarms that cost your operators the most time
- Your label schema and file format
- Enterprise licences already include one custom pack of up to 200 simulations, delivered within 6 weeks
We train our own models on this data.
Who we are
Xentim is the video analytics brand of TATO STUDIO sp. z o.o., a team from Poznan, Poland, that has worked professionally with images for years, and for the last few with artificial intelligence models that understand video.
Why we do it
Detectors learn from millions of photos of people and cars, but they almost never see smoke at night, a fall on a platform or a fight at an entrance. We needed that footage for our own models, so we started making it.
What are synthetic CCTV datasets?
Synthetic CCTV datasets are training data made in a game engine instead of recorded on real cameras. Xentim simulates falls, fights, railway track intrusions and fire and smoke, then exports the labels with every frame, so computer vision teams can train detectors without collecting footage of real people.
Which Xentim dataset should I start with?
Start with the event you cannot film. Fall Detection Dataset for person-down alarms, Violence Detection Dataset for fights and look-alikes such as hugs, Railway Track Intrusion Dataset for platforms and metro. Each has a free sample and a datasheet.
Do the datasets include video?
No. The fall, fight and rail packs contain skeleton data and labels, with no RGB images or video. The clips on this page are preview renders from the same simulations. The Fire and Smoke Starter pack is video.
Do you build custom datasets?
Yes. Send us your scene, camera layout, the events and look-alikes you need and your label format. We reply with the scope. Enterprise licences include one custom scenario pack of up to 200 simulations, delivered within 6 weeks.
How is a Xentim dataset different from public datasets?
Public datasets are mostly real footage that is small or restricted to research. Xentim packs hold thousands of simulated events with per-frame labels, 12 cameras per event, a steady 30 frames per second and a commercial licence. The comparison table on each dataset page cites its sources.
Is the data real?
No. Our datasets are synthetic and their datasheets say so. They contain no personal data of real individuals, so there are no people to get consent from.
Can I use it in a commercial product?
Yes, with a Standard or Enterprise licence. Research editions are non-commercial. Each dataset page links to the full licence text.
Can I test it before buying?
Yes. Each dataset has a free sample and a datasheet, so you can check the format and labels on your own model first.
How do I order?
Choose a licence and send us your order. We reply with an order form and invoice, you pay by bank transfer, SEPA or card, and we send a secure download link within 3 business days of payment.
What if the data does not match the datasheet?
Tell us within 30 days of delivery. We correct the dataset or, if we cannot, refund the fee against deletion of all copies.
Who can buy?
Companies and institutions. Research editions are for academic and public research institutions, and we review each application before delivery.
Do prices include VAT?
No, prices are net. Polish buyers pay 23% VAT. EU companies with a valid VAT ID pay under reverse charge. You receive an invoice with your order.
What does the licence not allow?
Our licences ban using the data for generative AI. Further limits differ by dataset, so read the licence linked on each dataset page.
Test it on your model first.
The sample is free. If the data does not match its datasheet, you have 30 days to tell us.