kiselyovd/citysample-vehicle-keypoints-24pt

City Sample Vehicle Keypoints (24-point, synthetic)

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).

Used by: kiselyovd/citysample-vehicle-keypoints-24pt - a YOLO-pose model trained entirely on this data.

Contents

• 1,440 frames at 1280x720 PNG, across 4 downtown venues x 3 lighting presets x 4 vehicle types (police sedan, taxi, and others).
• 15,390 annotations (multi-instance: the rig vehicle plus every visible city vehicle per frame), annotations/coco.json in COCO keypoint format.
• captures_all.jsonl - the raw per-frame records the COCO is built from.

Usage

from datasets import load_dataset

ds = load_dataset("kiselyovd/citysample-vehicle-keypoints-24pt")
ex = ds["train"][0]
ex["image"]                 # PIL.Image, 1280x720
ex["objects"]["bbox"]       # list of [x, y, w, h] (COCO), one per vehicle
ex["objects"]["keypoints"]  # list of flat [x1,y1,v1, ... x24,y24,v24] per vehicle

Each row is one frame; objects holds every labelled vehicle in it. Keypoints are a flat 24 x (x, y, v) list in the order in The 24-point schema below; v is 2 visible / 1 self-occluded / 0 absent.

Repository layout (two formats, same data)

The dataset ships in two interchangeable formats so it works with both ecosystems:

• data/*.parquet - the canonical Hugging Face format. Powers the dataset viewer and load_dataset (images + keypoints embedded, streaming-friendly, queryable with DuckDB / Polars / Pandas). Use this for research and HF tooling.
• g**/rgb/*.png + annotations/coco.json + captures_all.jsonl - the same frames and labels as raw image files in COCO-keypoint format. Use this for Ultralytics / YOLO and other pipelines that read image files and COCO directly (this is how the consumer model was trained).

Both describe the identical 1,440 frames; pick whichever your pipeline expects.

Splits

Split	Frames
train	1,152
validation	144
test	144

Stratified 80/10/10 by group (every venue x lighting x vehicle appears in all three splits). Orbit frames within a group are correlated, so for a strict leakage-free benchmark hold out whole groups instead - the group is the file_name prefix (e.g. g03_v1_day_clear). A raw COCO (annotations/coco.json) and the source captures_all.jsonl are also in the repo for COCO-style tooling.

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.

Exact index order (the flat keypoints list is [x, y, v] per index):

0  Right_Front_wheel        8  Exhaust                  16 Front_Left_Bumper_Corner
1  Left_Front_wheel         9  Right_Front_Top          17 Front_Right_Bumper_Corner
2  Right_Back_wheel         10 Left_Front_Top           18 Rear_Left_Bumper_Corner
3  Left_Back_wheel          11 Right_Back_Top           19 Rear_Right_Bumper_Corner
4  Right_Front_HeadLight    12 Left_Back_Top            20 Windshield_Bottom_Left
5  Left_Front_HeadLight     13 Center                   21 Windshield_Bottom_Right
6  Right_Back_HeadLight     14 Left_Side_Mirror         22 Rear_Window_Bottom_Left
7  Left_Back_HeadLight      15 Right_Side_Mirror        23 Rear_Window_Bottom_Right

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.

Honest status

This slice was built to test a kill switch: does synthetic pre-training improve a real-world model? Results are reported transparently in the source repo - a narrow single-city slice is a hard case for sim-to-real transfer, and negative or marginal results are documented rather than hidden. Use it as a research artifact and a pipeline demonstration.

License & provenance

Frames are non-interactive media rendered with Unreal Engine from City Sample. Under the UE EULA, such rendered images are freely distributable; no Epic asset files are included - only rendered PNGs and JSON annotations. The MIT license covers these frames and annotations, not Epic's underlying assets. See the EULA analysis.