Replacing DCL footage with clean synthetic data

Synthetic LED-lit gate dataset.

DCL The Game footage produces noisy labels and the wrong visual signature for the AIGP simulator due in May 2026. These synthetic frames render LED-lit square gates against procedural venue backgrounds with pixel-perfect labels — bbox + 4 corners — and AIGP-spec camera parameters.

Generator

synth_aigp_gates.py

pure CV2 + numpy, no GPU needed

Camera

640×360 · fx=fy=320 · 20° upward tilt

VADR-TS-002 §3.8 · HFoV=90° / VFoV≈58.7°

Gate

1.5 m square LED-lit, multi-color

white core + colored halo

Labels

YOLO bbox + YOLO-pose 4-corner

drop-in for Phase 1 + 2

What changed. Renderer uses a 5-layer LED bloom (dark structural frame · wide outer halo · mid bloom · inner glow · white-tinted hot core) so frames read as actual emitters, not flat outlines. Distance distribution beta-skewed toward 4–15 m (typical racing band). Backgrounds simplified to dark venue gradients with pinpoint distant LEDs — no random rectangles cluttering the frame.

§ 01Preview gallery

50 sample images at deploy/documents/synth-samples/preview/. A representative subset below — full set is shipped with the docs site.

§ 02How the renderer works

Stage	What it does
1. Random pose	Beta-skewed distance (3–25 m, mode≈8 m), azimuth/elevation within FOV, 6-DoF rotation that tightens at close range.
2. Background	Dark vertical gradient (15% chance bright/daylight), faint horizon line, a sprinkling of distant pinpoint LEDs with their own halos.
3. Project corners	Pinhole projection from 4 gate corners (1.5 m square in gate-local frame) → image pixels via fx,fy from FOV.
4. Frame body	Dark polyline (8,8,10) at 1.6× LED thickness — the matte structural part of the gate.
5. LED bloom (3 layers)	Big halo (×0.55, σ≈big), mid bloom (×0.95, σ≈med), inner glow (×1.4, σ≈small) — additive composite onto the scene.
6. Hot core	70% white + 30% LED color polyline, slightly thinner than the bloom — the actual emitter line.
7. Augmentations	Motion blur (60% chance), exposure jitter, hue shift, sensor noise, JPEG compression at random quality 45–95.
8. Labels	YOLO bbox (axis-aligned over corners) + YOLO-pose (4 corners with visibility flag).

Why white core + colored halo. Real LEDs are saturated outside, near-white at the source — that's what makes the eye read "emitter" instead of "drawn outline". Two-layer trick: bloom is the colored glow, core is white-tinted, both are additive over the dark frame.

§ 03Generate the dataset

Quick preview (50 images, ~10 s)

python synth_aigp_gates.py --output dataset_gates_synthetic --preview 50

Outputs JPEGs under dataset_gates_synthetic/preview/. Inspect, tweak, re-run.

Full overnight dataset (10K images, ~5–10 min)

python synth_aigp_gates.py --output dataset_gates_synthetic \
  --n-train 8000 --n-val 2000 --imgsz 640 360 --seed 42

Outputs the YOLO-format tree:

dataset_gates_synthetic/
  images/{train,val}/*.jpg
  labels/{train,val}/*.txt           # YOLO-pose: bbox + 4 keypoints (Phase 2)
  labels_bbox/{train,val}/*.txt      # plain YOLO bbox (Phase 1)
  data.yaml                          # for detector training
  data_pose.yaml                     # for keypoint training

Hook into existing training

Phase 1 (detector):

python train_apex.py detector \
  --dataset dataset_gates_synthetic \
  --epochs 200 \
  --name apex_yolo11n_synth \
  --wandb-project aigp-gate-detector-synth

Phase 2 (keypoints) reads data_pose.yaml automatically when the dataset has a kpt_shape field.

§ 04Known limitations & next steps

Procedural backgrounds, not photorealistic. Good enough for sim2real with augmentation, but a HDRI library (real venue panoramas) would push generalization further. Add --bg-dir arg and sample from there.
No occlusion modeling. Real races have other drones, posts, banners, judges occluding gates. Add: random dark rectangles overlapping gate corners with prob 0.2.
Single rectangular shape. AIGP spec doesn't pin down whether gates are square, rectangular, or hexagonal — when the spec lands, swap gate_corners_local().
No depth/distance label. Keypoints + bbox cover Phase 1 + 2. If we add a depth-regression head later, the renderer already has the gate's z in camera frame — log it.
Motion blur is linear-only. Real FPV racing has rotational blur during yaws. Add radial blur kernel for tight-turn variants.

Cross-refs: W&B uploads · training runbook · related code: synth_aigp_gates.py, train_apex.py:790 (camera constants).