weather-whiplash/trackside-condition

Trackside Track-Condition Dataset

Frame-level racing-surface condition labels for motorsport imagery: Dry / Damp / Wet.

Built for the Weather Whiplash project - a live track-condition detector that tells a pit wall whether the circuit is getting safer or riskier, and therefore when to change tyres.

Why this dataset exists

The nearest existing resource is RSCD (Road Surface Classification Dataset), an autonomous-driving dataset of dry / wet / water road patches. It is a useful signal but sits in a different visual domain:

	RSCD	This dataset
Camera	bumper-height, car-mounted	trackside / broadcast / onboard
Framing	close-up road patch, surface fills the frame	wide, surface is part of a scene
Surface	public asphalt	racing asphalt, painted kerbs, run-off, gravel traps
Confounders	traffic, lane markings	spray plumes, tyre marbles, a drying racing line

A classifier trained only on car-perspective road patches has to generalise across all four rows at once. This dataset exists to close that gap.

Classes

Class	Definition
Dry	Uniform pale-grey asphalt, no sheen
Damp	Uniformly darkened asphalt, no standing water, no spray
Wet	Standing water, strong reflections, visible spray

There is no Drying class, and its absence is a finding rather than an oversight. Of the 61 images CLIP assigned to "drying", none of the 25 checked were drying tracks: they were aerial circuit maps and sunny street scenes. With no verified examples the class was dropped automatically.

Drying is a property of a sequence, not a frame. A damp track and a drying track can look identical in one image; the difference is the direction of travel over the previous few frames. The project derives it from a trend layer over consecutive predictions instead.

Sourcing and licensing

Every image comes from Wikimedia Commons, primarily the Formula One in rain in the <decade>s category tree plus wet race weekends and circuit categories. Broadcast screenshots were avoided because they cannot be redistributed, which would make this dataset unpublishable.

Per-image title, author, licence and source URL are in ATTRIBUTION.csv. Licences are CC-BY / CC-BY-SA / public domain; the aggregate is released as CC-BY-SA-4.0. If you use this dataset, retain the attribution file.

Labelling method

1. Candidate pool collected from Commons categories (weak category-level prior).
2. CLIP zero-shot triage into the four classes plus a reject bucket (paddock portraits, garages, crowds - anything where the surface is not visible).
3. Human review of contact sheets, correcting the triage. Corrections override CLIP.

Step 3 is what stops this being a distillation of CLIP. Hand-verified images are allocated to the test and validation splits first, so reported accuracy is measured against checked labels.

Splits

Split	Dry	Damp	Wet	Total
train	141	34	169	344
validation	29	6	36	71
test	29	6	36	71

Total images: 486
Human-reviewed: 222
CLIP triage corrected by a human on 48.6% of reviewed images.

Limitations

• Skewed toward Formula One and European circuits; other series and surfaces are thinly represented.
• Damp is the hardest class for annotators and for models. Its boundaries with both Dry and Wet are genuinely continuous, and it is by far the least represented, because photographers shoot dramatic conditions and a merely damp track is not dramatic.
• Photographs are composed shots, not uniformly sampled video frames, so the distribution is not identical to a real fixed trackside camera feed.