Nothing is thrown away at capture
The engine writes RGB, the per-instance identity buffer, float depth in metres, a per-frame record of every actor and the camera, and a spawn manifest. Labels are computed from those afterwards.
Formats
Export is a conversion, and every conversion loses something.
A capture writes a raw dump first: frames, the identity buffer, depth in metres, and a record of where everything was. YOLO and COCO are derived from that afterwards. Both are narrower than the source, which is fine as long as you know which parts they leave behind.
- Exporters
- COCO, YOLO
- Source
- raw dump, always kept
- Reference capture
- 40 frames
- Instances
- 4,378
The raw dump is the real format
Everything else is a projection of it, which is why the dump is what a run keeps.
Re-export without re-rendering
Because derivation happens outside the engine, the same capture becomes a YOLO set, a COCO set, or a format that does not exist yet, without opening the editor again.
Both ship together
The sample packs carry the exported labels and the buffers they came from, so you can check one against the other.
If you need a format that is not listed here, it is usually a small converter over the dump rather than a change to the capture. Tell us which one.