Until the official AIGP simulator drops, DCL The Game is the closest visual match to VQ2 we can get — DCL operates AIGP's technical side and the Steam game uses real 3D-scanned tracks with FPV camera views. Play the game, capture_dcl.py runs in the background recording + auto-labeling frames, and you wake up with thousands of VQ2-style training images in YOLO format.
dataset_gates_dcl/dataset_gates_mega (2.7K generic drone-racing images) gets 97.9% mAP50 on its own val set. VQ2 will be visually very different — realistic 3D-scanned environments, harder lighting, distractors. DCL gameplay is the highest-ROI source of VQ2-style training data available before the AIGP sim releases.
store.steampowered.com/app/964570 · ~$20 · Windows native.
Fullscreen borderless also works but windowed is easiest for our capture tool to locate. In-game graphics settings → Display → Windowed, resolution 1920×1080 recommended.
The capture tool uses mss (screen capture) and pygetwindow (locate game window by title). Both are installed in the aigp venv. If you recreate the venv:
./aigp/Scripts/python.exe -m pip install mss pygetwindow# Default: locates a window with "DCL" in title, uses current APEX detector
./aigp/Scripts/python.exe capture_dcl.py
# If auto-detect misses, pass the exact title substring
./aigp/Scripts/python.exe capture_dcl.py --window "DCL The Game"
# Last resort: fixed screen region (left, top, width, height)
./aigp/Scripts/python.exe capture_dcl.py --region 0 0 1920 1080
# Progress counter while you play (run in yet another terminal)
./aigp/Scripts/python.exe capture_dcl.py statsCtrl+C in the capture terminal stops the session cleanly. Data is saved incrementally — nothing is lost if you crash out.
| Flag | Default | What it does |
|---|---|---|
--window | DCL | Substring match on game window title |
--region | — | Capture a fixed rect instead of a window |
--model | models/apex_yolo11n.pt | Detector used for auto-labeling |
--fps | 10 | Capture rate in Hz (higher = more frames, more dup reject) |
--conf | 0.40 | Only save frames where detector confidence ≥ this |
The capture script only saves a frame if it survives three filters:
| # | Filter | Why |
|---|---|---|
| 1 | Dark-frame reject mean brightness < 25 discarded | Menu screens, loading screens, fade-to-black are garbage for training. |
| 2 | Near-duplicate reject dHash hamming distance < 3 discarded | If you hover before a gate for 5 seconds, you'd otherwise get 50 near-identical frames wasting disk and biasing the loss. |
| 3 | No-gate reject confidence < 0.40 discarded | Frames with no visible gate add nothing. Hard negatives are captured separately (different pipeline). |
Everything that passes is split 85% train / 15% val deterministically (hash on frame index) so you can rerun safely without re-splitting.
| Dimension | What to do | Why |
|---|---|---|
| Tracks | Fly ≥ 3 different DCL tracks per session | Different environments (stadium, industrial, outdoor) = visual diversity |
| Speed | Mix scenic laps + race-pace | Covers both sharp detail and motion-blurred frames |
| Angles | Deliberately pitch up / down / sideways | Off-axis gate views that race laps don't capture |
| Lighting | Play day and night tracks if available | VQ2 spec mentions "dynamic lighting" |
| Failures | Crash on purpose near gates sometimes | Partial occlusion, extreme angles, post-impact frames |
| Distance | Approach gates slow from far away | Small-gate detection (5-30 px) is what fails at 200m approach |
./aigp/Scripts/python.exe capture_dcl.py statsExpect ~1-2K frames per hour of varied gameplay. Less than that = either the detector is rejecting too much (lower --conf) or capture-region is wrong.
Open dataset_gates_dcl/images/train/ in File Explorer. Pick 20 random files. For each, open the matching dataset_gates_dcl/labels/train/<name>.txt and mentally check the box coords overlay correctly. YOLO format is class cx cy w h (all normalized [0,1]).
audit_dcl.pyAuto-labeling gets you ~80% clean data. The last 20% — the obvious detector misses, duplicate hover frames that slipped through dedup, and gates that shouldn't have been labeled — needs a human eye. audit_dcl.py is a keyboard-driven batch reviewer that makes this take 2-3 sec per frame.
# Review everything (resumes from last stop)
./aigp/Scripts/python.exe audit_dcl.py
# Only review frames with ≥ 2 boxes (multi-gate shots are noisier)
./aigp/Scripts/python.exe audit_dcl.py --min-boxes 2
# Print progress
./aigp/Scripts/python.exe audit_dcl.py stats
# Restore a rejected pair if you changed your mind
./aigp/Scripts/python.exe audit_dcl.py --restore dcl_20260424_120000_001234| Key | Action |
|---|---|
Enter / → / k | KEEP — advance |
Del / ← / d | REJECT — move to rejected/ (recoverable), advance |
u | UNDO last action |
Space / s | SKIP — no decision, advance |
b | BACK — previous file |
1..9 | jump N × 10 files ahead |
h | toggle help overlay |
q / Esc | QUIT (progress auto-saved) |
Rejections are moved to dataset_gates_dcl/rejected/ — never deleted — so mistakes are recoverable via --restore. State persists in .audit_state.json so you can quit mid-audit and resume later.
The easy way: add DCL data as a separate phase in train_apex.py detector by editing the dataset arg:
# Train on mega first, then fine-tune on DCL
./aigp/Scripts/python.exe train_apex.py detector --dataset dataset_gates_mega --epochs 200
./aigp/Scripts/python.exe train_apex.py detector --dataset dataset_gates_dcl --epochs 50The smarter way: merge both datasets into a single YOLO config via symlinks or a combined data.yaml pointing at both roots. Pattern already used in dataset_gates_domain / dataset_gates_hardneg.
| Pitfall | Symptom | Fix |
|---|---|---|
| Window not found | "No window matching 'DCL'" | Script prints all open windows. Pass the exact substring with --window. |
| 0 frames saved | stats shows 0 images after 5 min | Detector conf threshold too high (default 0.40). Drop to 0.25: --conf 0.25. |
| Too many near-identical frames | Dataset full of hover shots | Play more, hover less. Or tighten dHash threshold (edit script line _hamming(h, last_hash) < 3 → < 5). |
| GPU spiking during game | Game stutters | Detector inference uses ~1-2 GB VRAM. On RTX 5080 + DCL this should be fine. If issues, lower --fps 5. |
| DCL fullscreen-exclusive | mss captures desktop not game | Switch to Windowed or Fullscreen-Borderless in DCL's display settings. |