Drom.ru: Computer Vision and ML Price Millions of Vehicles and Detect Body Damage from Photos
Computer vision automatically verifies and tags millions of uploaded vehicle photos every month, slashing manual moderation workload while auto-filling make, model, and generation for sellers.
| Company & Sector | Дром (Drom.ru / FarPost) — Auto Classifieds & Marketplace |
|---|---|
| Models & AI Stack | Computer Vision (CNN / ViT) · License Plate OCR · Visual Embedding Search · Gradient Boosting Valuation · Fraud & Duplicate Detection · Vehicle History NLP |
| Rollout Timeline | 2018–2020 — launch of neural make/model recognition from photos and license plate OCR; 2021–2023 — production ML fair-market valuation and visual duplicate search; 2024–2026 — multimodal moderation and LLM summarization of owner reviews. |
| Verification | Verified 2026-10 · 3 primary sources |
Business Context & Scale
Drom.ru (part of the FarPost technology group) is one of Russia's leading automotive marketplaces, serving over 20 million monthly unique users across millions of vehicle listings and auto parts catalogs.
Used-car transactions involve high ticket values and severe information asymmetry. Prices for the exact same make, model, and year can vary by 30–50% depending on region, trim, drivetrain, mileage, and body condition.
Drom's machine learning team pioneered neural vehicle recognition from user photos alongside automated ML fair-market valuation across every listing on the platform.
Business Problem & Bottlenecks
In automotive classifieds, sellers frequently mislabel model generations or facelifts, upload low-light photos, obscure license plates, or reuse stolen photos to post bait-and-switch scam listings.
Second is fair price estimation. Naive averages by make and year fail badly: right-hand-drive vs. left-hand-drive imports across regions, hybrid vs. naturally aspirated engines, and accident history create massive price spreads.
Third is deduplication and anti-fraud: detecting commercial flippers and scammers who repost identical vehicles using cropped or mirrored photos.
AI Architecture & Implementation
When a seller uploads photos on Drom, a Computer Vision cascade (CNNs and Vision Transformers) automatically classifies camera angles (exterior front/side, interior cabin, engine bay, title document), identifies make, model, and body generation with 95%+ accuracy, reads and masks license plates for vehicle-history cross-checks, and flags visible body damage.
To catch duplicate and stolen photos, perceptual and deep visual embeddings search across tens of millions of historical listing images in milliseconds, surviving crops, color filters, and horizontal flips.
Drom's fair-price valuation engine ('Great / Fair / High Price' badges) runs gradient boosting over hundreds of signals: regional closing velocity, exact trim, ownership/accident records, and visual condition embeddings.
“"A vehicle photo on Drom carries as much signal as the seller's form: the neural network reads the exact body generation, trim details, and repair clues while cross-checking the image against millions of historical listings." — Drom Engineering
Measurable Impact & Economics
Computer vision automatically verifies and tags millions of uploaded vehicle photos every month, slashing manual moderation workload while auto-filling make, model, and generation for sellers.
The ML valuation engine covers the vast majority of passenger-car listings with industry-leading low MAPE error, giving buyers instant price transparency and flagging suspiciously underpriced scam listings.
Visual embedding deduplication significantly reduced fraudulent reposts and boosted buyer trust in Drom vehicle history reports.
Key Takeaways for Industry
In auto marketplaces, Computer Vision pays off three ways at once: seller autofill, stolen-photo fraud detection, and license-plate history linking.
Used-car pricing models must explicitly capture sharp regional market divides and actual listing-to-sale velocity.
Severe negative deviation from the ML fair-price estimate is one of the strongest signals for bait-and-switch scam detection.
Millions of long-form owner reviews and logbooks provide a proprietary grounding corpus for AI reliability summaries.
Verified Primary Sources
Key Questions Answered
How does the vision model distinguish pre-facelift from facelift vehicle generations?
Trained on millions of labeled catalog photos on Drom, the network focuses on headlight geometry, grille patterns, bumper contours, and dashboard layouts.
Why doesn't the 'Great Price' badge simply go to the cheapest listing?
Extreme low outliers often indicate salvage vehicles, title issues, or fraud. The valuation model incorporates VIN history and condition signals while filtering suspicious anomalies.
What if a scammer flips or crops a stolen car photo?
Deep visual embeddings are invariant to horizontal flips, minor rotations, and cropping, linking the image back to the original historical listing in milliseconds.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.