When a technician corrects a missed dent, a false positive, or a wrong dent size, is that correction captured as labeled training data?

Yes. Every technician decision on a verified panel is stored as a label. A model detection the technician keeps is a true positive, one they remove is a false positive (it is soft-deleted, never erased, precisely so it stays a labeled example), a dent they add is a miss the model made, and a size the technician changes is a re-bucketing. Those four labels are what Estimoto's precision, recall, dent-count error and size-confusion metrics are computed from.

Do corrections improve our shop's own model, or a shared model?

Both, and the private half is the one you feel first. Your shop's corrections fit a calibration used only for your shop: a confidence floor and size-class shifts learned from your own verified panels, re-fitted after every verification and applied to your next panel. The same panels also train the shared V-JEPA head, with a cap on how many images any single shop contributes so no one shop dominates. The Learning page shows exactly what your technicians have taught detection and how much of your shared-model cap you have used. Gold panels are never used to fit either half.

Does Estimoto store the original image, the AI detections and the technician's corrections?

Yes, all three, on every panel. The original photograph is kept as the source of truth. Each AI detection is stored with its position, diameter in millimeters, size class, confidence score, which photograph found it, and how many angles corroborated it. Technician adds, removes and resizes are stored against those detections, and a manually added dent carries the technician's measured diameter and a close-up photograph.

Can I see AI confidence per dent?

Every detection carries a confidence score from 0 to 1 and a corroboration count (how many photographs of the panel saw the same dent). Both are returned on the detection through the API. The phone app does not display the number today; it uses corroboration to decide what to show the technician.

Can I export detections and corrections?

Yes. An owner or manager can export every per-dent label the shop owns as CSV from the Learning page: the original image, the AI's position, size class and confidence, how many angles corroborated it, the technician's verdict (kept, removed or added), any size correction, who verified it and the vehicle it was on. Estimates also export as an insurer-ready PDF and as CIECA-BMS into CCC ONE, Mitchell or Audatex.

Can managers compare the technician's read against the AI's read and against what the carrier finally approved?

Yes, in two views. The Learning page reports detection accuracy per technician: precision, recall, F1, dent-count error and size accuracy for the panels each technician verified, next to the shop total and the held-out gold panels. A technician sees their own row; owners and managers see everyone. Separately, each technician has a scorecard of estimates sent, approved and denied and what carriers cut from their work, so the AI read, the technician's read and the carrier's answer sit side by side.

Does it classify dent size?

Yes, into four classes: dime, nickel, quarter and half-dollar, each with a millimeter threshold (half-dollar is 27 mm and up). There is no separate oversized class; a dent a technician adds by hand carries its measured diameter in millimeters, so anything larger than a half-dollar is recorded by size rather than lost in the top bucket.

What are your independently measured precision and recall?

Estimoto does not publish figures yet; it publishes the bar instead. Detection today is served by Claude vision (Anthropic's Opus 5 model). Estimoto's own V-JEPA dent-detection head is trained on technician-verified panels and is allowed to replace Claude only when, on held-out gold-standard panels, it reaches precision of at least 0.70, recall of at least 0.60, an F1 of at least 0.65 and size-class accuracy of at least 0.70, and does not fall below the model it replaces on any of them. Precision and recall are always computed separately from technician decisions, never collapsed into one accuracy number, and gold panels are never used for training. Measured figures for whichever model is serving will appear on this page, with the method and the panel count, once at least eight gold-standard panels exist to score against.

How does performance change on white, black, silver, metallic, dirty, wet or reflective panels?

Accuracy is reported per vehicle colour and per paint finish (solid, metallic, pearl or matte) from the colour and finish recorded on the job, so a shop can see how detection does on its black metallics versus its solid whites. Wet or dirty panels are not tagged today, so those are not sliced yet, and the page will say so until they are.

Is a reflection or light board required?

No. Capture is an ordinary phone camera with guided angles per panel. No light board, reflection board or rig is required.

Can multiple technicians have individual accounts and reporting?

Yes. Each technician signs in with their own account and role. Every estimate is attributed to the technician who wrote it, and the technician scorecard reports sent, approved and denied counts, approval rate and carrier cuts per person.

Is there a senior-technician validation step before a correction becomes ground truth?

Yes, in two places. A technician can mark a verified panel as a gold standard; a gold mark from anyone outside a small trusted set goes to a review queue and changes nothing until approved. Separately, a training-only candidate-review queue lets a reviewer assess model predictions without touching any live estimate. What is not measured yet is repeatability, whether the same technician reads the same panel the same way twice.

Paintless dent repair

AI dent detection for PDR technicians

Estimoto detects hail and door dents from photographs taken on an ordinary mobile phone, then turns them into a priced, carrier-aware estimate. Technicians correct what the models get wrong, and those corrections train Estimoto's own vision model, so detection improves on the vehicles your shop actually works on.

Get a photo estimate How hail estimating works

Capture, detect, correct, verify, learn

Dent detection is not a single model call. It is a loop that closes on the technician, and every pass through it makes the next read better.

Step 1 of 5: Capture
Photograph the damage where the car sits
A technician walks the vehicle with an ordinary phone. Estimoto guides the capture panel by panel and asks for the angles a dent read actually depends on, so the light rakes across the panel instead of flattening it. There is no booth, no scanner and no rig to move the car onto. A shop can capture in a driveway, a hail lane or a customer's parking lot, and a customer can capture their own vehicle from a link the shop texts them.
Step 2 of 5: Detect
Three models read every panel at once
Each panel is analyzed by V-JEPA, OpenCV and Claude vision together, and their reads are fused rather than ranked. Estimoto returns a dent count and a size distribution per panel, which is what a PDR matrix prices from. Cross-angle fusion reconciles the same dent seen from two photographs so it is counted once, not twice.
Step 3 of 5: Correct
The technician has the final say
Every detected dent is editable. A technician can add a dent the models missed, remove one they invented, and resize any dent whose severity was read wrong. Nothing is locked. The estimate that goes out is the one the technician signed off on, which is the only version worth learning from.
Step 4 of 5: Verify
A panel is marked verified, and gold when it is exemplary
When a technician confirms a panel is right, that panel becomes verified evidence. The strongest panels can be marked gold standards, which are deliberately held OUT of training and used to grade the model instead, so accuracy is measured against work no model was allowed to learn from. Gold marks from anyone outside a small trusted set go to review before they count.
Step 5 of 5: Learn
Verified panels improve the model over time
Confirmations, removals and resizes all feed a dedicated V-JEPA head that keeps training on your technicians' corrections. Accuracy is measured against your own technicians rather than a generic public benchmark: every confirmed, removed or missed dent moves precision, recall and panel-level error. The system gets sharper on the vehicles and the hail your shop actually sees.

How accuracy is measured

Estimoto grades itself against panels its models were never trained on. Technicians mark exemplary panels as gold standards; those are held out of training and used only for scoring, which is what keeps the number honest. Precision, recall and panel-level dent-count error are tracked per panel type and per vehicle. Detection is served by Claude vision today; Estimoto's own V-JEPA head replaces it only when it clears precision 0.70, recall 0.60, F1 0.65 and size-class accuracy 0.70 on those held-out gold panels, and beats the model it replaces on each one. Those floors are the public bar, and they do not move to let a candidate through.

Published benchmark figures will appear on this page as that held-out evidence set grows. Estimoto would rather show a number drawn from real technician-verified panels than a marketing claim, so until there is enough of it, the honest answer is that accuracy is measured against your own technicians and reported back to you.

The questions to ask any dent detection vendor

These are the questions a shop should put to every product it evaluates, answered for Estimoto against the code that runs in production. Where the honest answer is "not yet", it says not yet.

When a technician corrects a missed dent, a false positive, or a wrong dent size, is that correction captured as labeled training data?: Yes. Every technician decision on a verified panel is stored as a label. A model detection the technician keeps is a true positive, one they remove is a false positive (it is soft-deleted, never erased, precisely so it stays a labeled example), a dent they add is a miss the model made, and a size the technician changes is a re-bucketing. Those four labels are what Estimoto's precision, recall, dent-count error and size-confusion metrics are computed from.
Do corrections improve our shop's own model, or a shared model?: Both, and the private half is the one you feel first. Your shop's corrections fit a calibration used only for your shop: a confidence floor and size-class shifts learned from your own verified panels, re-fitted after every verification and applied to your next panel. The same panels also train the shared V-JEPA head, with a cap on how many images any single shop contributes so no one shop dominates. The Learning page shows exactly what your technicians have taught detection and how much of your shared-model cap you have used. Gold panels are never used to fit either half.
Does Estimoto store the original image, the AI detections and the technician's corrections?: Yes, all three, on every panel. The original photograph is kept as the source of truth. Each AI detection is stored with its position, diameter in millimeters, size class, confidence score, which photograph found it, and how many angles corroborated it. Technician adds, removes and resizes are stored against those detections, and a manually added dent carries the technician's measured diameter and a close-up photograph.
Can I see AI confidence per dent?: Every detection carries a confidence score from 0 to 1 and a corroboration count (how many photographs of the panel saw the same dent). Both are returned on the detection through the API. The phone app does not display the number today; it uses corroboration to decide what to show the technician.
Can I export detections and corrections?: Yes. An owner or manager can export every per-dent label the shop owns as CSV from the Learning page: the original image, the AI's position, size class and confidence, how many angles corroborated it, the technician's verdict (kept, removed or added), any size correction, who verified it and the vehicle it was on. Estimates also export as an insurer-ready PDF and as CIECA-BMS into CCC ONE, Mitchell or Audatex.
Can managers compare the technician's read against the AI's read and against what the carrier finally approved?: Yes, in two views. The Learning page reports detection accuracy per technician: precision, recall, F1, dent-count error and size accuracy for the panels each technician verified, next to the shop total and the held-out gold panels. A technician sees their own row; owners and managers see everyone. Separately, each technician has a scorecard of estimates sent, approved and denied and what carriers cut from their work, so the AI read, the technician's read and the carrier's answer sit side by side.
Does it classify dent size?: Yes, into four classes: dime, nickel, quarter and half-dollar, each with a millimeter threshold (half-dollar is 27 mm and up). There is no separate oversized class; a dent a technician adds by hand carries its measured diameter in millimeters, so anything larger than a half-dollar is recorded by size rather than lost in the top bucket.
What are your independently measured precision and recall?: Estimoto does not publish figures yet; it publishes the bar instead. Detection today is served by Claude vision (Anthropic's Opus 5 model). Estimoto's own V-JEPA dent-detection head is trained on technician-verified panels and is allowed to replace Claude only when, on held-out gold-standard panels, it reaches precision of at least 0.70, recall of at least 0.60, an F1 of at least 0.65 and size-class accuracy of at least 0.70, and does not fall below the model it replaces on any of them. Precision and recall are always computed separately from technician decisions, never collapsed into one accuracy number, and gold panels are never used for training. Measured figures for whichever model is serving will appear on this page, with the method and the panel count, once at least eight gold-standard panels exist to score against.
How does performance change on white, black, silver, metallic, dirty, wet or reflective panels?: Accuracy is reported per vehicle colour and per paint finish (solid, metallic, pearl or matte) from the colour and finish recorded on the job, so a shop can see how detection does on its black metallics versus its solid whites. Wet or dirty panels are not tagged today, so those are not sliced yet, and the page will say so until they are.
Is a reflection or light board required?: No. Capture is an ordinary phone camera with guided angles per panel. No light board, reflection board or rig is required.
Can multiple technicians have individual accounts and reporting?: Yes. Each technician signs in with their own account and role. Every estimate is attributed to the technician who wrote it, and the technician scorecard reports sent, approved and denied counts, approval rate and carrier cuts per person.
Is there a senior-technician validation step before a correction becomes ground truth?: Yes, in two places. A technician can mark a verified panel as a gold standard; a gold mark from anyone outside a small trusted set goes to a review queue and changes nothing until approved. Separately, a training-only candidate-review queue lets a reviewer assess model predictions without touching any live estimate. What is not measured yet is repeatability, whether the same technician reads the same panel the same way twice.

Note on the name: Estimoto.io is the hail and collision estimating product for body shops described on this page. It is unrelated to the European motorcycle marketplace or the older mileage-tracking app that share the name.

What the detection feeds

Detection is the front of a longer workflow. The same verified damage drives the estimate, the carrier negotiation and the shop's day-to-day operations.

PDR and collision estimating

The same capture produces a priced estimate. Estimoto prices off carrier PDR and hail matrices, or a custom shop matrix where no carrier matrix exists, and recommends supporting operations such as R&I, R&R, teardown, pre and post scans, ADAS calibration and paint labor when a panel routes away from PDR. Every line is editable, and estimates export by CIECA-BMS or PDF into CCC ONE, Mitchell or Audatex.

Remote estimates

A shop can send a link by text or email and let the customer photograph their own vehicle. The estimate comes back without the car ever being driven in, which is what makes catastrophe and out-of-area work practical.

Rejection and approval coaching

Estimoto keeps the record of what each carrier denied or cut, why they said so, what the technician changed in response and what was ultimately approved. Before an estimate goes out it flags the lines that carrier has cut before, so the documentation is attached the first time instead of argued for later.

Multi-location CRM

Customers, job pipeline, calling and texting, parts ordering, rentals and scheduling, with visibility across every location an owner operates and per-location drill-down.

Agent automation for shop operations

Agents connect to Gmail, Outlook, Slack, Google Drive, Sheets, Docs and Calendar, plus QuickBooks and Mailchimp. They chase parts that are running late, flag overdue invoices, follow up with customers and brief owners on activity. Anything customer-facing waits for approval before it sends.

A clear next step

Choose the setup that matches your shop.

Choose the estimating app or CRM separately
See the price and workflow before checkout
Add the second product only when your shop needs it

Choose your product Sign in to your dashboard

AI dent detection for PDR technicians

Capture, detect, correct, verify, learn

Photograph the damage where the car sits

Three models read every panel at once

The technician has the final say

A panel is marked verified, and gold when it is exemplary

Verified panels improve the model over time