rdo1/euro-dashcam-vehicle-dataset

European Dashcam Vehicle Crops

STATUS: RAW / UNVERIFIED — vehicle crops with automatically-generated labels. A provider-verification pass (DeepSeek/Grok/Gemini) is in progress; a CLEANED subset with double-confirmed labels will be published as verified/ when ready.

Honest label-quality disclosure (important)

• —Labels were generated automatically (DeepSeek vision, batch 50/request)
• —Independent sanity check (Grok, 300-crop sample): ~55% agreement — most disagreements are confusable near-identical models (Audi Q3↔Q5, Tesla Model 3↔Y, Renault Clio↔Citroën C3, Ford Focus↔Toyota Prius), not random noise
• —Crops are small vehicle boxes (100-500 px) — distant/blurry cars are inherently hard
• —Use with caution for training/eval: expect ~45% of labels to be imperfect for hard confusable pairs. Cross-verify before trusting specific labels.

Provenance

• —Crops derived from publicly available YouTube dashcam footage of European cities (Berlin, Paris, Rome, Vienna, Prague, Barcelona, Munich, London, Istanbul, and more)
• —Non-commercial research use (CC-BY-NC-4.0); original videos belong to their uploaders
• —No readable license plates included

Content

• —~21,300 crops across ~820 classes (make + model; generations merged at training)
• —Structure: one folder per class, JPG vehicle crops

How it was built

1. 1.yt-dlp → 1080p dashcam videos of European cities
2. 2.YOLOv8 + ByteTrack → one sharp crop per car per distance bucket
3. 3.DeepSeek batch labeling → dashcam_raw/<Label>/
4. 4.Gemini/Grok second-opinion recovery of low-confidence rejects
5. 5.build_merged.py: accent-fold, dedupe, 5x oversample, leak-free split

Companion model

rdo1/euro-dashcam-vehicle-classifier — evaluated honestly on a permanent frozen holdout (crops no model ever trained on): see the model card for real numbers.