Modelpublic

weather-whiplash/vit-track-condition

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes17downloads

Model Card

ViT - Trackside Track-Condition Classifier

Classifies a racing-surface frame as Dry / Damp / Drying / Wet.

Fine-tuned from google/vit-base-patch16-224-in21k on trackside motorsport imagery collected from Wikimedia Commons. Built for Weather Whiplash, a live track-condition detector for race strategy.

Usage

python

from transformers import pipeline

pipe = pipeline("image-classification", model="weather-whiplash/vit-track-condition")
pipe("trackside_frame.jpg")

Why fine-tune instead of zero-shot

CLIP zero-shot with surface-worded prompts is a reasonable baseline and is kept in the project as a fallback, but it confuses an overcast sky with a wet surface and cannot represent Drying (a dry racing line on wet asphalt) at all - there is no natural-language prompt that reliably isolates it.

Evaluation

—validation: accuracy 0.831, macro-F1 0.698
—test: accuracy 0.831, macro-F1 0.747

Per-class on the held-out test split:

Class	Precision	Recall	F1	Support
Damp	0.40	0.67	0.50	6
Dry	0.92	0.79	0.85	29
Wet	0.89	0.89	0.89	36

The test split is hand-verified, so these numbers are measured against human labels rather than triage output.

Intended use and limits

Decision support for race strategy, not an autonomous decision-maker. A single frame is never enough to call "drying" - that requires the trend layer in the project repo, which reads direction of change across a rolling window of frames.

Training data skews to Formula One and European circuits. Expect degradation on night races, heavy motion blur, unusual surfaces (street circuits, gravel) and other series.

Training data

See the companion dataset. Images are CC-BY / CC-BY-SA from Wikimedia Commons with per-image attribution retained.