Keypoints and pose
The capture contract has room for it and the plugin can carry pose-enabled actors, but the reference capture has none and nothing here is validated for it. It will get a page when there is a run to point at.
Tasks
What you can actually train on this.
A capture is not a dataset for every task at once. These are the three a NameFrame run supports properly today, with the ground truth each one gets and what it costs. Anything not listed here is either derivable from the raw dump with work, or not supported, and this page says which.
- Reference capture
- 40 frames
- Classes
- 6
- Instances
- 4,378
- Object detectionAxis-aligned boxes with class labels, computed from the visible pixels of each object rather than from projected bounds.YOLO and COCO
- Instance segmentationPer-object masks from a dedicated identity buffer, with occlusion resolved by the renderer and fragments reassembled.COCO polygons, raw ID buffer
- Depth estimationDense float32 metres per pixel, unnormalised, alongside the frame it belongs to.NumPy arrays
What each task gets from one frame
All three come from the same capture. You do not choose a task at capture time and lose the others.
| Task | Ground truth | Derived from | In the sample pack |
|---|---|---|---|
| Object detection | Class + box | identity buffer | yes, YOLO and COCO |
| Instance segmentation | Class + per-object mask | identity buffer | yes, COCO polygons |
| Semantic segmentation | Class per pixel | identity buffer | derivable from the ID buffer |
| Depth estimation | float32 metres | depth buffer | yes, 3 frames |
Semantic segmentation does not have a page of its own under tasks because it is a projection of the instance masks rather than a separate output. How that derivation works is worth reading if you need it.
Not supported today
Tracking across frames
Objects are reshuffled every frame by design, so consecutive frames are independent samples rather than a sequence. A tracking dataset needs a different capture mode.
Optical flow, normals, materials
The engine has all of them. The capture does not currently write them, so they are not claimed here.