kiselyovd/citysample-vehicle-keypoints-24pt

City Sample Vehicle Keypoints (24-point, synthetic)

Capture explorer · Generator · Transfer experiment

Live 14-point real-data detector is a separate model/application; it does not run the synthetic-only 24-point weights.

A synthetic vehicle-keypoint dataset rendered inside Epic's City Sample ("Matrix Awakens") with Unreal Engine 5.6 and Movie Render Queue. Vehicles are placed on the real ZoneGraph road network, and every visible vehicle is labelled with a 24-point anatomical keypoint schema plus a mesh-bounds bounding box.

Generated by the open pipeline at kiselyovd/ue5-vehicle-synth (full method, code, and engineering write-up in the repo wiki).

Source & capture workflow · Transfer experiment

Contents

• 2,510 frames at 1280x720 PNG, across 36 capture groups: 4 city-street venues x 3 lighting presets (clear day, golden hour, overcast) x 3 hero vehicle rigs (sedan, minivan, pickup). Each scene is additionally populated with parked car/van types drawn at random from the City Sample fleet, so labelled instances cover more body shapes than the three hero rigs.
• 14,286 annotations (multi-instance: the rig vehicle plus every visible labelled city vehicle per frame, ~5.7/frame), annotations/coco.json in COCO keypoint format.
• Per-point visibility split: 55.9% visible (v=2), 34.2% occluded (v=1), 9.9% off-frame/absent (v=0) - the capture is deliberately biased toward hard framing and occlusion.
• captures_all.jsonl - the raw per-frame records the COCO is built from.

The 24-point schema

Points 0-13 are the CarFusion canonical order (4 wheels, 4 head/tail lights, exhaust, 4 roof corners, body center) for backward compatibility with 14-point models. Points 14-23 extend it: 2 side mirrors, 4 bumper corners, 4 window-base corners. Per-point visibility follows CarFusion: 2 visible, 1 self-occluded, 0 off-frame.

Every keypoint location is derived geometrically from the vehicle's own mesh data - wheels from the skeletal wheel bones, roof corners from the body vertex cloud's roofline plateau, head/tail lights and mirrors from the centroids of the mesh sections bound to the light/mirror materials - so per-type configs are exact to each body shape (a van's high tail lights, a pickup's cab-only roof), not a scaled sedan template.

Intended use

Synthetic pre-training / augmentation for real-world vehicle-keypoint models (sim-to-real). It is a Phase-0 vertical slice: the focus is the pipeline and a proof-of-concept sample, not a final large-scale corpus.

Used by

• Synthetic 24-pt model - a YOLO-pose model trained entirely on this synthetic data (the in-domain model).
• Real CarFusion model - the real-world baseline the sim-to-real study compares against.

Evaluation status

The Phase 0 matched experiment is complete. Adding synthetic frames did not meet the +2 pp OKS-mAP criterion on held-out CarFusion: the supplement contributed -1.7 pp from the weak checkpoint and -9.5 pp from the strong one. This result applies to the tested 2D transfer recipe, independently of the synthetic-only in-domain benchmark. Read the protocol and error analysis.

License & provenance

Frames are non-interactive media rendered with Unreal Engine from City Sample. The published artifact contains rendered PNGs and JSON annotations; it does not include Epic asset files. Its Hub license metadata remains MIT for these published outputs. Epic's underlying City Sample assets are outside that metadata and retain their own terms. See the EULA analysis.

Training split and published partitions

Actual pinned Parquet membership has now been compared with the source COCO and original checkpoint training lists: 234 of 257 Hub test frames were training frames for the released checkpoint. All 2,510 IDs/paths/dimensions and 14,286 annotations were checked; image bytes were not compared. The model used 2,259 train / 251 validation frames; Hub publishes 1,996/257/257. The Hub test is not an independent test for these weights.

Hub split	Checkpoint train	Checkpoint validation
train	1,792	204
validation	233	24
test	234	23

Exact crosswalk and comparison receipt. Crosswalk SHA256: 840a9f538f95eed38889e2cdd4ab5ac97928d183a469bd57175ec980b92fcd0c. Related orbit views also cross splits; this does not establish scene/vehicle independence.