Salvage auction data — photos, damage labels, and what happened to the car next

Three signals, and none of them is a hammer price. The repair-cost estimate on the lot; the pre-loss retail value (what the car was worth before the damage); and the actual asking prices from retail listings where the same VIN reappeared after the auction. That last one is the number most models want anyway — it is what the vehicle was really offered for once it was back on the road. We do not carry the winning bid for a lot and we do not model one: salvage auction outcomes are not published in a form we can verify record by record, and we will not sell a price field we cannot stand behind.

Copart and IAAI, across the United States, with the auction city and state on every record so you can filter or stratify by geography. The observation window runs from June 2021 to the present, model years reach from the late 1990s to the current year, and about 97% of lots are model year 2000 or newer. New lots are added daily.

They originate in the auction houses' own intake records and are normalized by us into about 30 consistent classes — front end, rear end, side, hail, water/flood, rollover, burn, vandalism, mechanical, minor dents and scratches, undercarriage, all over, and others. The labels describe the vehicle, not a specific region of a specific photo: there are no bounding boxes or masks in the standard delivery. Part-level annotation is available as a paid add-on if your model needs it.

For about 561,000 VINs we have matched the salvage auction record to the retail listings that same VIN appeared in afterwards. You get whether it came back, how many days later, and the asking prices, mileage, seller, and location of each listing — a market event rather than an estimate. It is partial by construction: a VIN appears only if the car genuinely returned to the retail market and we observed the listing. That biases the linked subset toward lighter damage and dealer resale, and we would rather you model around it than discover it in month two. Two things make that tractable. First, reappeared_on_retail_market: false means “not observed”, not “never resold” — so the honest framing is positive-unlabeled learning, not binary classification. Second, every auction-side covariate is present on both the linked and the unlinked lots, so you can estimate the propensity to be observed and reweight. Ask us for the linkage rate broken down by damage class and we will send it with the sample.

Yes. Licences cover commercial AI and machine learning training, and are available period-limited or perpetual, per dataset or combined. If you intend to redistribute or white-label the data rather than train on it, that is a different arrangement — the redistribution page linked in the delivery section covers it. Some sources carry attribution requirements your account team will identify during scoping. Personal information is stripped before delivery and a data processing agreement is available on request.

The sample package is a ZIP of 20 complete lots — 285 photographs, one record.json per lot, and a README with the data dictionary and licence terms. In the last package we sent, 18 of the 20 lots carried a post-auction retail outcome. Tell us the segment — makes, model years, damage classes, states — and we filter it to match, before any agreement is signed. Turnaround is typically about one business day and the sample is evaluation-only. Full delivery is CSV, Parquet, or JSON to your S3 bucket, Snowflake, BigQuery, or SFTP endpoint, or through the API, with images as signed URLs.

Public research sets — CarDD, VehiDE and the Kaggle car-damage packs — run from roughly two thousand to fifteen thousand images and are annotated by academic teams for machine learning research, often with per-region masks. This is a commercial corpus of 528M+ photographs across 46.2M+ salvage auction lots, with ~44M carrying an auction-house damage class and ~4.4M also carrying a repair-cost estimate and a pre-loss retail value on the same record. The trade-off is granularity: our labels are lot-level unless you add part-level annotation as a paid add-on, and in exchange you get scale, an economic label, and a post-auction retail outcome that no public set carries.

For about 561,000 VINs, yes. We match the salvage auction record to retail listings the same VIN appeared in afterwards, so you get whether the car came back, how many days later, and the asking price, mileage, seller and location of each listing. The 2023 Kia Rio on this page was back in retail listings twenty days after its Copart record, asked at $15,659 against a $16,875 pre-loss retail value and a $1,626 repair estimate. The hail-damaged 2025 Toyota Corolla took 144 days and was listed at $28,520. Those are asking prices, and the linkage is partial by construction.