Context
The sim generates per-camera 2D projections + visibility, which is exactly what a COCO/YOLO training annotation needs. Exporting to those formats turns the sim into a synthetic training-data generator, not just a benchmark.
What to do
Add a typed exporter that reads a SceneManifest and writes COCO (json) and YOLO (txt) 2D bbox/keypoint annotations per frame per camera, using entity ids as categories and in_view/visible for the visibility flag. Round-trip test: export then parse back, boxes match the manifest uv within tolerance.
Context
The sim generates per-camera 2D projections + visibility, which is exactly what a COCO/YOLO training annotation needs. Exporting to those formats turns the sim into a synthetic training-data generator, not just a benchmark.
What to do
Add a typed exporter that reads a SceneManifest and writes COCO (json) and YOLO (txt) 2D bbox/keypoint annotations per frame per camera, using entity ids as categories and in_view/visible for the visibility flag. Round-trip test: export then parse back, boxes match the manifest uv within tolerance.