tobil/racing-gears

Racing Gear Digits

Image classification dataset for detecting gear numbers (0-9) from racing onboard camera telemetry overlays, supplemented with MNIST digits for robustness.

Similar in spirit to MNIST but for a specific real-world application: reading the gear indicator from racing car onboard video feeds in real-time.

Dataset

• 5,964 training / 1,003 validation images
• 32×32 grayscale PNG images
• 10 classes (digits 0-9)
• Source column distinguishes racing-original, paul-ricard-alpine, sebring-tobi-lap6, racing_aug (augmented), and mnist
• Proper stratified split — 15% of each racing source held out for validation
• Augmented minority classes — racing digits with <200 training samples augmented via random shifts, brightness/contrast jitter, and Gaussian noise

Racing sources

Source	Gears	Style
TDS Racing IMSA Sebring 2026 (original)	1-6	White digit on gray RPM gauge face
Sebring Q Tobi Lap 6	1-6	White digit on dark semi-transparent overlay
Paul Ricard Alpine LMPh	1-7	White digit on dark circle
MNIST supplement	0-9	Handwritten digits (generalization)

Train distribution

Digit	Racing (orig)	Sebring	Paul Ricard	Aug	MNIST	Total
0	0	0	0	0	196	196
1	864	36	35	0	196	1,167
2	1,344	74	105	0	196	1,719
3	364	99	162	0	196	821
4	37	68	4	91	196	396
5	107	138	49	0	196	490
6	106	79	35	0	196	416
7	19	0	45	137	196	397
8	0	0	0	0	196	196
9	0	0	0	0	196	196

Adding new video sources

1. Extract gear crops from a video:

   uv run python scripts/extract.py <video_path> <source_name> <x> <y> <w> <h>
   

   This creates raw/<source>/unlabeled/ frames and a composites/<source>/unlabeled.png contact sheet.

2. Label by reading the contact sheet and creating labels/<source>.csv:

   start,end,label
   0,14,5
   15,39,6
   

   Each row maps a frame range (inclusive, 0-indexed) to a gear digit.

3. Build the dataset:

   uv run python scripts/build_dataset.py
   

   This reads all raw/ sources and labels/ CSVs, does stratified train/val splitting, augments minority classes, and writes the parquet files.

Augmentations (racing_aug)

For racing classes with fewer than 200 training samples, synthetic samples are generated:

• Random translation — up to ±3px shift in x/y
• Brightness jitter — 0.7–1.3×
• Contrast jitter — 0.8–1.2×
• Gaussian noise — σ=8, 30% probability

Usage

from datasets import load_dataset

ds = load_dataset("tobil/racing-gears")

# Filter to real racing images only
racing = ds["train"].filter(lambda x: x["source"] not in ("mnist", "racing_aug"))

# Standard training loop
for example in ds["train"]:
    image = example["image"]   # PIL Image, 32x32 grayscale
    label = example["label"]   # int 0-9
    source = example["source"] # source identifier

Context

Built for the mpv racing telemetry plugin which reads baked-in telemetry from racing onboard videos using mpv's screenshot-raw API and renders a live overlay with throttle/brake traces, gear indicator, and steering position.

The gear digit is detected using a small CNN (ONNX) called via LuaJIT FFI from mpv's Lua scripting environment at 30fps.