Generate It, Annotate It, Ship It: Building CCTV-Ready Training Data Without Touching a Labeling Tool

Labeling is the part of the computer vision job nobody puts on a slide. Generating the model is fun. Training is fun. Deploying is, on a good day, fun. Sitting in a labeling tool at 11pm dragging boxes around the same parking-lot frame for the fourteenth time is not fun, and it's where most CV projects quietly lose two months.

We just shipped the feature that removes that step entirely. You write a prompt — say, "a person walking through a public space carrying a small handheld pistol down by their side, full-body shot, CCTV camera" — Simuletic Studio renders the image, and a second later it hands you back YOLO-format bounding boxes for whatever classes you asked for. No labeling tool. No second vendor. No third invoice.

The screenshot at the top of this post is the actual output of that workflow — two frames, three classes, ready to drop into an Ultralytics training run.

The Two-Step Workflow That Used to Take Weeks

Here's what the Studio looks like with auto-annotation switched on. You pick a camera view (standard, CCTV, aerial / drone), write your scene, choose resolution, batch size, and the classes you want detected. Confidence threshold sits at 0.40 by default — high enough to filter the obvious junk, low enough not to miss small or partially occluded objects.

The Studio with auto-annotation enabled. Add up to eight classes per generation; we'll detect every instance up to ten boxes per class.

Hit generate. That's it. Each image gets a .txt file in standard YOLO format next to it (class id, normalized cx, cy, w, h) and the classes are saved as a data.yaml for you. The bounding boxes are toggled on and off with a single switch on the results page so you can sanity-check before exporting.

What's Actually Happening Under the Hood

We run a SAM3-based open-vocabulary detector over every generated frame, one class at a time so class identity is never lost in a mask batch. For each class we get back pixel-space x, y, width, height for every instance, then normalize, aggregate, and write a standard YOLO label file.

Two small things are worth knowing:

• Per-class detection cap of 10. Enough to cover a crowded parking lot or a dozen pedestrians without ever returning a wall of noise from a low-confidence model that thinks every shadow is a person.
• Tunable confidence (0.05 → 0.90). Default 0.40 is the sweet spot for medium-distance CCTV. Drop to 0.20 if you're shooting aerial / small-object scenes. Push to 0.60 if you want only the cleanest training signal.

CCTV Use Cases Customers Are Shipping In a Day

The first wave of teams using this are all building CCTV-adjacent systems. A few patterns we keep seeing:

"But Are the Boxes Any Good?"

Reasonable question. Look at the parking-lot frames at the top of this post: every car in both lots is boxed cleanly, including the partially-occluded vehicles in the back row. Both walking pedestrians are tight, full-body boxes. The handheld pistol — a 30-pixel-wide object inside a 1024-pixel image — is also caught in both frames.

The honest answer is: SAM3 is excellent at common, recognizable objects (people, vehicles, animals, large props), very good at moderately specific terms (weapons, drones, backpacks, knives), and where it needs help is on small objects in busy scenes — that's where dropping the confidence to 0.25–0.30 pulls them back in. The threshold is right there in the panel, no API call needed.

For teams who need pixel-perfect labels on niche classes, this is still a 95%-of-the-way tool: it gives you a high-quality first pass that a junior reviewer can clean up in minutes per image instead of hours.

Export Straight to YOLO

Hit export on any dataset and you get a single zip: images/, labels/ with one .txt per image, and a data.yaml with your class names. Drop it straight into Ultralytics, Roboflow, or your own training pipeline. No conversion script. No "wait, what format is this in" Slack thread on a Friday afternoon.

Try It

Auto-annotation is live in Studio for every plan, including the free tier. Pricing is the part that's almost embarrassing: 0.5 credits per class actually detected in an image. One credit if you asked for person and weapon and we found both. Half a credit if only one class hits. You pay for results, not attempts.

The best labeling tool is the one you never have to open. Generate the scene, get the boxes, train the model. That's the whole loop now.

Related reading: Shoplifting detection dataset, Smoke & fire detection, and Weapon detection.