AGENTS.md

Agent Context: Racing Gear Digits Dataset

Domain knowledge

• A lap in circuit racing takes ~1.5–2.5 minutes depending on track length. During a single lap the driver uses every gear at least once (except reverse/neutral).
• So any ~2 minute window of onboard footage guarantees examples of all forward gears (typically 1–6 or 1–7 depending on car).
• Gear changes are not uniform — the car spends much more time in mid-range gears (2, 3, 4) than in 1st or top gear. Low gears appear briefly during braking zones; top gear only on long straights.
• The gear indicator is part of the telemetry overlay baked into the video. Its position, font, size, and background vary per video source. Each source needs its own crop coordinates.
• Gears 0, 8, 9 don't occur in normal racing — 0 is neutral (rare), 8+ don't exist on these cars. We include them via MNIST for classifier completeness.

Video sources

Each video source has different crop coordinates for the gear digit:

Video	Resolution	Gear crop (x, y, w, h)	Gears	Notes
Sebring Q Tobi Lap 6	1920×1080	(1440, 780, 90, 105)	1–6	White digit on dark semi-transparent overlay
Paul Ricard Alpine LMPh	832×464	(685, 320, 55, 55)	1–7	White digit on dark circle

Working with the data

• Use uv run for all Python scripts (dependencies managed via pyproject.toml)
• Extraction scripts live in scripts/, reference examples live in examples/
• The dataset parquet files are in data/ and tracked with Git LFS
• Source column values: racing (real crops), racing_aug (augmented), mnist (handwritten)