Objects a robot must not hit
Ground truth for grasping, navigation and obstacle avoidance, in layouts you can re-roll thousands of times. Every object's position is known in metres, so depth and pose supervision come free with the frame.
Use cases
Where this is worth the trouble.
Synthetic data is not free. You pay for it in scene work, and you pay again in the gap between a render and the world. It is worth paying when the data you need is dangerous to collect, rare by design, or does not exist yet. These are the problems we think NameFrame fits, stated as hypotheses because that is what they are.
Aerial views nobody wants to fly for
Altitude, pitch and lens are parameters instead of a flight plan. The airbase capture on this site is exactly this: a camera 60 to 190 metres out, looking down at a scene too dense to annotate by hand.
Defects and states that are rare on purpose
The failure you need a thousand examples of is the one your process is designed never to produce. In a scene, the rare case is just another configuration.
Clutter, occlusion and stacking
Pallets behind pallets, people between racks. Occlusion is handled by the identity buffer rather than estimated, so a partly hidden object still gets the box it deserves.
One viewpoint, every condition
A camera bolted to a wall sees one scene under every light, weather and season it will ever face. Those conditions can be generated in an afternoon instead of collected across a year.
Controlled variables
One thing changes and everything else holds still. Determinism makes an ablation an ablation rather than two runs that happened to differ.
What we are not going to tell you.
Every synthetic data company on the internet has a page of industry logos and a claimed percentage improvement. We are in Commercial Preview with a small number of teams, so we do not have that evidence yet, and inventing it would be the fastest way to lose the argument with anyone who has actually trained a model.
No ROI numbers
We have not measured a return for your industry, so there is no percentage here. When a pilot produces a number we are allowed to publish, it will appear with the method that produced it.
No claimed domain gap
How far a synthetic set gets you depends on your task, your real data and your model. The tooling to measure that gap ships in the product; the answer for your case is something you get by running it.
What we can show
A real capture, its distributions, its failing check, and a download. Read the quality page →