AIGP · PQ TRAINING DATASET GEN YOLO11n-POSE · 4-CORNER KEYPOINTS

Generates synthetic training frames by flying a camera along randomized race courses — now across four VQ1-style topologies (A2RL hangar · long-straight · speed-loop · technical) for realistic variety in gate spacing, scale, and multi-gate perspective. Matches your existing train_apex.py dataset format exactly (nc=1, kpt_shape=[4,3], names=[gate]). Output is a ZIP with train/val split, data.yaml, and a README — drop into dataset_gates_synthetic_pose_pose/ and run training.

Generation parameters

unique layouts
samples per course
% held out
82 = A2RL default
leave blank = Date.now()
0 / 0 — fps · ETA —
DATASET READY
—
total frames
0
labeled gates
0
train / val
0 / 0
avg gates/frame
0.00

Spec compliance · VADR-TS-002

Resolution 640 × 360 px · §3.8
Intrinsics fx=fy=320 · cx=320 · cy=180
FoV H 90° · V 58.715°
Tilt +20° up · NED · pinhole · §3.8
Format JPEG · UDP:5600 30 Hz · §4.6
Gate 1500 mm inner · 2700 mm outer · 260 mm deep · §3.7
Appearance consistent across track · single-class · §3.1
Venue 100 × 30 m × 15 m hangar · ADNEC Marina Hall scale
Courses A2RL · long-straight · speed-loop · technical

Output structure

aigp-pq-dataset.zip
├── images/
│   ├── train/synth_train_NNNNNN.jpg
│   └── val/synth_val_NNNNNN.jpg
├── labels/
│   ├── train/synth_train_NNNNNN.txt
│   └── val/synth_val_NNNNNN.txt
├── data.yaml
├── manifest.json
└── README.md

Preview · last frame

course— frame— t— gates labeled— size—

Top: synthetic frame · the JPEG the AI will process · all gates uniform blue per spec (no color cues).
Bottom: same frame with YOLO labels rendered for verification — bbox + 4 corner keypoints per gate.