timdwms/us-car-models-1948-2026

555 US Car Models, 1948-2026

222,832 freely licensed photos of 555 car models from Wikimedia Commons, sorted into folders by make, model and generation, like Toyota/RAV4/XA50 (2019-2025)/. Every photo comes with its labels, the author's name, its license, and a link back to its Commons page.

I collected and labeled these for my Vehicle Identifier project, a model that names a car's make, model, generation and facelift from a photo. I'm sharing them so other people learning machine learning have a car dataset they can actually use and redistribute.

At a glance

• 222,832 photos, 49 makes, 555 models, 1,425 generations
• Cars from 1948 to 2026, mostly 1990 onward: models sold in the US, photographed around the world
• 189,515 passed a quality filter (a clear, single car that fills the photo); the rest are included too, with the reason they were dropped
• 150,548 have their exact look (generation + pre-facelift or facelift) labeled
• Photos are 800 px wide JPEGs (resized from the originals on Commons), about 39 GB in total

Not every car

This is not every car ever made, or even every car sold in the US. It covers:

• 49 makes and 555 models: the popular US-market models, chosen for the Vehicle Identifier project. Brands that were never sold in the US (Peugeot, Renault, ...) and models sold only in other countries aren't included.
• 1,425 generations, from the first Ford F-Series pickups (1948) to 2026 models. Most photos are of cars from 1990 onward.

For a sense of scale: the US government's vehicle database (NHTSA vPIC) lists 1,413 model names for these brands from 1990 to 2026, and this dataset matches 701 of them (50%). Many of the rest are trims and body styles of a model that is included (Audi A8 L, BMW 740e), commercial trucks (Ford F-590, Volvo NE42) or rare models (Subaru XT6, Geo Storm).

What's in it

metadata.csv              one row per photo (see below)
images/Toyota.zip         unzips to Toyota/<Model>/<Generation (years)>/<photo>.jpg
images/Ford.zip
...                       one zip per make (49)
images/_Multiple cars.zip photos whose Commons categories name more than one car

The photos are zipped because Hugging Face works badly with hundreds of thousands of loose files. Unzip the makes you want and you get plain folders.

Two special folders:

• <Make>/<Model>/_Unknown generation/ (1,495 photos): the Commons category only names the model, not which generation.
• _Multiple cars/ (7,695 photos): the photo is filed under categories of different cars, for example two cars side by side, so there's no single label.

metadata.csv

Column	Meaning
file_name	Path inside the zip, e.g. BMW/X3/G01 (2018-2024)/8c6e3a67424b1a2d10b9.jpg
zip	Which zip it's in
make, model, generation, years	The car. years covers the whole generation.
phase	pre or facelift for generations with a mid-life update, single for generations without one, blank if unknown
look_label	How phase was decided (see below)
filter	ok if the photo passed the quality filter, otherwise why not: no clear vehicle, vehicle too small, two vehicles, duplicate
detector_score	How sure the car detector was that there's a vehicle (0-1)
car_box	The main car's box as left top right bottom, in pixels of the 800 px image
used_in_run6	train, val or test if my model's 6th training run used this photo, blank if not
commons_category	The Commons categories the photo was downloaded from
commons_title, commons_url	The photo's name and page on Wikimedia Commons
author, license, license_url	Who made the photo and its license (see Licenses)
date_taken	When the photo was taken, if Commons knows
original_width, original_height	Size of the full-resolution original on Commons

How the labels were made

All labels were made by scripts, not checked one by one by a person. They're good, but expect some mistakes.

1. Make, model and generation come from the photo's Commons category ("Toyota Corolla (E140)"), matched to a hand-built table of about 2,100 car generations and facelifts.
2. Quality filter: a pretrained car detector (Faster R-CNN) checked every photo. It rejects photos with no clear vehicle (interiors, engines, badges), a car that's too small, two similar-sized cars, or an exact duplicate of a photo already kept.
3. Pre-facelift vs facelift (look_label), for photos that passed the filter:

look_label	Photos	How
auto: category	109,286	The category only fits one look
auto: year in filename	23,053	A model year in the file name picks the look
auto: model prediction	18,089	My trained model was clearly sure which look it is
human	120	Picked by hand
needs review: ...	38,967	The clues disagreed, or my model didn't know the look yet. The make, model and generation are still from the category.
blank	33,317	Not sorted, mostly because the filter dropped the photo

A label was only given automatically when independent clues agreed (the category, the year in the file name, the date the photo was taken, and the model's prediction). Photos where they disagreed are marked needs review.

Licenses

There is no single license for this dataset. Each photo keeps its own license, set by its author on Wikimedia Commons. All of them allow reuse and sharing:

License	Photos
CC BY-SA (versions 1.0 to 4.0, including country versions)	133,581
CC BY (versions 1.0 to 4.0, including country versions)	51,438
Public domain	21,955
CC0	15,858

The license column has each photo's exact license and version, e.g. cc-by-sa-4.0.

If you use or share these photos:

• Credit the author for CC BY and CC BY-SA photos: the author, license and commons_url columns have everything you need.
• CC BY-SA photos must stay CC BY-SA if you share them or changed versions of them.
• The photos here are resized to 800 px wide. The originals are on each photo's Commons page.

The labels and metadata.csv (my work) are released under CC BY 4.0.

Some photos show people, license plates or logos. Commons allows these, but depending on where you are and what you use them for, other rights (privacy, trademarks) may apply.

Limitations

• Labels are machine-made and will contain some mistakes, especially in needs review rows.
• Uneven coverage: popular cars have hundreds of photos (capped at 200 per Commons category, or 800 for cars from 2015 on), rare ones only a few.
• Mostly US-market models, because the car table was built for US cars.
• Car show bias: many Commons photos are clean, well-lit shots from car shows, not everyday street photos.

Related

• Code, training pipeline and results: https://github.com/timmyd4/Vehicle_Identifier
• My model was also trained on VMMRdb (Tafazzoli et al., CVPR Workshops 2017). Those photos are not included here and can't be redistributed; download them from the official source.