NameFrame/airfield-crowd-unreal-synthetic

Airfield crowd: a synthetic pose capture rendered in Unreal Engine

300 rendered frames from a single Unreal Engine environment, carrying 4,469 labelled object instances. Every box and every mask here is read out of the engine's own per-instance ID buffer at render time. No model produced these labels and no one drew them by hand, so a label is wrong only where the scene description behind it is wrong.

The capture

Engine	5.8.1-56057345+++UE5+Release-5.8
Map	Map_Airbase_Demo
Frames	300
Instances	4,469
Resolution	1280 x 960
Horizontal FOV	60 deg
Camera policy	random, 2 authored zones
Lens profile	one of fifteen camera profiles per frame, 25 frames untouched
Time of day	08:30 to 17:30, across fourteen weather and time cells
Weather	Ultra Dynamic Sky presets: clear, partly cloudy, cloudy, overcast, foggy, light rain and rain
Master seed	20260929
Validator	grade A, 94.7 out of 100

Placement, pose and camera pose are sampled deterministically from the master seed (deterministic_per_logical_frame_and_instance), so the same seed over the same scene reproduces the same 300 frames.

Classes

id	class	instances
0	person	4,469

By COCO's scale thresholds: 2,490 small, 1,944 medium, 35 large. The mean box covers 0.10% of the frame, which makes this a small-object capture.

Layout

• images/ frames through one of fifteen camera profiles, drawn per frame: drone, CCTV and phone looks with grain, softness, haze, lens dirt and compression, at 0.82 to 1.18 of full strength. 25 of the 300 are left untouched. The profiles recolour pixels and move none, so every box, mask and keypoint lines up with the image as it does with the raw render
• segmentation/ per-instance ID buffers, one colour per instance
• labels/ YOLO-pose rows, one text file per frame: each box followed by 17 joints with a visibility flag, on people
• annotations/instances.json the same boxes in COCO format, without the keypoints
• metadata/ per-frame camera, actor and spawn-manifest records
• reports/ this capture's own report and validation output
• capture.json the run contract: classes, camera policy, seed
• data.yaml class names, ready for a YOLO trainer
• test.txt, train.txt, val.txt the export's split as image lists, which data.yaml points at

Frame N is the same scene across images/, segmentation/, labels/ and metadata/. The split ships with the pack: train.txt, val.txt and test.txt list the frames in each, and data.yaml points at them. It was chosen so that no cluster of near-identical frames straddles train and validation.

Decode segmentation/*.png with a library that hands you channels in the order the file declares. The instance colours are exact 8-bit values, and a silent channel swap turns a correct mask lookup into an empty one.

Measured quality

This archive ships the validator's own output rather than a summary of it. From reports/validation.json:

• Corrupt images: 0 of 300
• Frames with no labels: 0
• Near-duplicate images: 56
• Cross-split leakage: 0
• Mean luminance 86.1; 0 underexposed frames, 0 overexposed
• Reconciled frame by frame: 300 requested, 300 images, 300 label files, 4,469 annotations

Known limitations

• The keypoint rows have the 17 COCO slots and only the 12 body joints carry values. Nose, eyes and ears are 0 in every row, because the capture records one head point; mask those five out when training.
• The validator took 5.0 points off for near-duplicate-images: visually near-identical images in one split (dHash <= 4).
• The validator took 0.33 points off for blurry-images: variance-of-Laplacian < 100.0.
• Check build-provenance did not pass: BUILD_PROVENANCE_UNVERIFIED: UNVERIFIED_REPOSITORY_COMMIT.
• Depth-probed people the engine could see, against people labelled: 22/4176 engine-visible people unlabelled (0.53%, 300/300 frames).
• Cameras are drawn from two boxes over the apron and the service yard, 2.5 to 25 m up and 12 to 45 m from their target, so the same hangars and silos recur. This is a sample of one environment, not a benchmark.
• Rendered imagery carries a synthetic-to-real domain gap. Treat it as pre-training or augmentation material rather than a substitute for a real-world evaluation set.

Provenance and terms

The frames were rendered in Unreal Engine and labelled from the engine's instance buffers by NameFrameCapture, an editor plugin. capture.json carries the full build record, including the plugin and job hashes and the checks the build could not verify.

Environment and prop assets in the scene are licensed from third parties for use inside Unreal Engine projects. The rendered images and their labels are published here; the source assets are not, and no .uasset or map file is included. Read LICENSE before redistributing the imagery itself.

The capture pipeline, the other environments in this series and the measurements behind them are documented at https://getnameframe.com.