Skip to content

Published dataset and model split lineage

The published dataset at bf2e175ddfb0b2dcd8dee3264f38eec3c19e643b contains 1,996 train / 257 validation / 257 test frames. The synthetic-only checkpoint at bb0971f76e0944a3dd90faa236cbff37547b5590 was trained with a different 2,259 train / 251 validation split.

All 2,510 actual pinned Parquet IDs, filenames and dimensions were compared with the source COCO. All 14,286 annotations were compared at absolute tolerance 0.0005 for float32 bbox/keypoint values. Only annotation/identity columns were read, approximately 5 MiB; image-byte equality was not tested.

Published dataset split Existing checkpoint train Existing checkpoint validation
train 1,792 204
validation 233 24
test 234 23

The published test split is not independent of this checkpoint's training data. Do not evaluate the released model on Hub test and present the result as held-out generalisation. The reported synthetic validation results refer to the original 251-frame model validation fold. Frame splits also share related capture/orbit views, so they do not prove independence by scene or vehicle identity. This does not change the separate matched real-CarFusion transfer experiment.

Exact ID/path crosswalk · Parquet comparison receipt · Source/training identity

Crosswalk SHA256: 840a9f538f95eed38889e2cdd4ab5ac97928d183a469bd57175ec980b92fcd0c. The dataset and model splits remain unchanged; this receipt documents their relationship.

Validate or repeat

python scripts/validate_lineage.py verifies the tracked crosswalk and report, unique IDs, exact counts and actual training basename checksums. To repeat the remote column comparison, run scripts/verify_hub_lineage.py --coco PATH --training-images PATH --out PATH in the declared project environment. The original source COCO and training image lists are required; the script rejects a changed COCO and oversized remote reads. No full image-column download is needed.

Separate component terms

Code and author-created annotations are declared MIT. Rendered images incorporate Epic City Sample content. The Epic Content License Agreement separately addresses rendered linear media, UE-only content and other restrictions; the author's MIT declaration does not relicense Epic asset files. The dataset contains renders/annotations rather than source meshes or textures. Historical legal notes are the author's earlier interpretation, not a replacement for applicable upstream terms.