bounding-boxes

Use the project command with an existing RGB camera: it captures one frame, builds intrinsics from the camera fov and resolution, transforms each actor's 3D box vertices into camera space, and writes an annotated PNG plus a JSON of xyxy boxes.

Run the draw command with a filter such as vehicle.* to overlay green 3D boxes via world.debug.draw_box. The boxes are re-stamped each frame for the configured duration so they track moving actors, and the server must be rendered, not headless.

Projection works headless because it only reads a camera frame and actor poses, but the draw command produces no visible output on a -nullrhi server since debug overlays require rendering. Use project or list for headless annotation workflows.

No boxes appear when actors are behind the camera, beyond the --max-dist range, fully off-frame, or excluded by the filter. Raise --max-dist, widen the filter, and confirm vehicles are actually in the camera's view.

Misalignment happens when the captured frame and actor poses come from different ticks during fast motion in asynchronous mode. Capture in synchronous mode so the frame and poses belong to the same simulation tick.