Plate Review
Plate Review reads the license plate in each image you send it, accepts the reads it is confident about, and puts the rest in a review queue where your team confirms or corrects the text. Every read, every decision and every settings change is recorded, so you can show how each plate was read and who checked it.
What it's for
Any job where license plates in photos have to become trustworthy text: tolling, parking enforcement, gate and lot access, fleet yards, or investigations. Plate Review takes care of the reading. Your team only looks at the plates that need a person, and the system learns from every one they check.
Navigating here
Plate Review is an app your company admin turns on for your company. Once it's set up, open it from your company's apps. Its pages are:
- Home shows what needs attention: reads waiting for review, reads auto-accepted today (and how), failed images, the reader's status, whether each kind of auto-accept is on, and the accuracy measured by spot checks.
- Upload adds a new batch of images.
- Review is the review queue, one read at a time.
- Reads is a sortable, filterable table of every read. Click a row to open the read's detail.
- Audit searches the audit trail and exports it.
- Settings holds the auto-accept switches, accuracy targets, connected storage and the history of every settings change.
- Exports downloads your reads and your labeled data.
Key concepts
- Batch. A group of images added together, from an upload, a connected storage folder or a tool call.
- Image. One photo. An image can show no plate, one plate or several. Each image is Queued, Processing, Read, No plate or Failed. An image that can't be opened or read is marked Failed with a plain-language reason. It is never silently dropped.
- Read. One plate found in an image: a close-up crop of the plate, the text the reader saw, and how confident it is in each character.
- Read status. A read is Auto-accepted, Needs review, AI checking (waiting a few seconds for the AI's own read), Not a plate (AI) (filed by the AI), Confirmed, Corrected, Unreadable or Not a plate.
- Label. Every read a person confirms or corrects becomes a label. Labels are how Plate Review learns which reads it can trust.
Adding images
Upload
- Open New batch.
- Drag in images, or a zip of images. Common photo formats are supported, up to 25 MB per image.
- Give the batch a name and add it. Reading starts right away.
Connected storage
Plate Review can watch a folder in your own cloud storage bucket and pick up new images as they arrive. The images stay in your bucket and are read where they are, so nothing is copied out first.
- In Settings, open Storage sources and choose to add a new source.
- Enter the bucket and folder, and how often to check for new images.
- Create an access role in your cloud account that lets Autessa read that folder. The role must require the external ID shown on the page. That ID is unique to your company, so no one else can use your role.
- Save. Each new image is picked up exactly once, even if the folder is checked many times. Images added with names that sort earlier than ones already picked up are still found.
If a check fails (for example, the role no longer allows reading), the source shows the error in plain language, and the next check tries again.
From another system (tool)
Plate Review comes with a plate_read tool, so your other systems, agents and scripts can send images. Call it with a list of images, each given as an uploaded file, an HTTPS link or a storage location. You can also pass when the photo was taken, the camera and the lane. The tool returns the batch's ID and the IDs of the queued images. To check on them later, use the Plate Review query tools, or open the batch in the app. See Tools for how to call a tool.
The first batch after a quiet period takes a few minutes to start while the plate reader starts up. Home shows Starting plate reader… with the number of images waiting. Once it's running, batches are read in seconds.
How confidence and auto-accept work
For every plate, Plate Review uses an open-source plate reader. It reports how sure it is that it found a plate and how sure it is of each character. Plate Review combines these into one score per read.
Auto-accept is on from the start. There are three switches in Settings → Auto-accept and thresholds, all on by default; you can turn any of them off. Home shows whether each one is on. Turning any of them on or off is saved as a new settings version in the audit trail, with who changed it and when.
Accept when the reader and the AI agree
Every plate the reader can't accept on its own also gets read by a separate AI model, which looks only at the plate close-up. When the AI is sure and reads exactly the same text, the plate is accepted. If the AI reads something different, isn't sure, or doesn't answer, the plate goes to review, with the AI's reading shown as a hint. A plate waits for the AI for a few seconds at most, and is never left waiting.
- AI checks a minute. A limit on how many plates the AI reads each minute (600 by default). Past it, plates go straight to review instead of waiting.
- With this switch off, those plates go to review, and the AI's reading is still shown as a hint.
Let the AI file what isn't a plate
The detector sometimes boxes something that isn't a vehicle plate: a sign, a sticker, a house number. With this switch on, when the AI is sure a crop isn't a vehicle plate, it is filed as Not a plate without a person. A plate the AI can't read, or isn't sure about, still goes to a person. These decisions are never used to teach the reader.
Accept plates the reader alone is sure of
Scores alone aren't a promise, so Plate Review measures how often a score turns out to be right, using your own team's labels:
- Learning period. Even with this switch on, nothing is accepted by the reader alone until your team has checked enough reads to prove the target accuracy. The stricter the target, the more reads it takes: about 765 at 99.5%, or about 190 at 98%. Home shows Learning: N of M labels.
- Target accuracy. After that, Plate Review finds the lowest score at which reads have been right at least as often as your target, by default 99.5%. It counts a score as safe only when the evidence clearly supports it, not just on average. Reads at or above that score are accepted.
- It keeps learning. Every night, Plate Review recalculates the score threshold from all your labels so far. Each recalculation is saved as a new settings version, with the accuracy it measured and the share of reads it would accept. It never turns a switch on or off.
Spot checks
A sample of what was decided without a person is sent to review anyway, marked as a spot check: by default 5% of the AI's decisions (plates it agreed on and crops it filed as not a plate), and 1% of plates the reader accepted alone. Home shows the accuracy measured on the last 30 days of spot checks for each kind, so you can see how auto-accept is really doing.
The review queue
Review shows a page of plates at a time, 12 by default (6 or 24 in the page-size menu). Every plate counts as right unless you mark it. Glance over the page, fix the few that are wrong, then confirm the whole page at once. The next page is already loaded, so there's no wait.
Each plate shows a close-up, the text the reader saw (characters it was unsure of are highlighted), the AI's reading when there is one, and a small copy of the photo with the plate boxed. The plates most likely to need you, such as an AI reading that differs, come first on the page.
Plates on your page are held for you, so two reviewers never work on the same plate. If you leave, they go back to the queue.
Batch review shortcuts
| Key | Action |
|---|---|
| Arrow keys (or h j k l) | Move between plates on the page. |
| Type | Correct the selected plate. Enter or Tab saves and moves on; Esc puts the reader's text back. |
| Alt+U | Mark the selected plate Unreadable. |
| Alt+N | Mark the selected plate Not a plate. |
| Alt+R | Clear your change to the selected plate. |
| Ctrl+Enter | Confirm the page: every plate you didn't mark is confirmed. |
Undo last page is available for 60 seconds after you confirm a page.
Review one by one
For hard cases, choose Review one by one. It shows one plate with the full photo beside it. The plate this card is about is outlined in yellow and tagged This plate. When a photo has more than one plate, the card says which one it is (for example, Plate 2 of 2 in this photo).
| Key | Action |
|---|---|
| Enter | Confirm the text as shown, or save your correction. |
| E, or click the text | Start correcting. Esc puts the reader's text back. |
| U | Mark the plate Unreadable. |
| N | Mark it Not a plate. |
| S | Skip it for now. Skipped plates aren't shown again this visit; when only skipped plates are left, you can start over with them. |
| Z, or click the photo | Zoom in on the plate. |
| ← / → | Step back and forward through the plates you've just reviewed. |
While you're typing a correction, use Alt+U, Alt+N and Alt+S instead.
When a photo has more than one plate
Check the yellow box: a card can be about a small or distant plate rather than the obvious one. If you type text that already belongs to another plate in the same photo, Plate Review asks before saving. Usually the right choice is Unreadable or Not a plate for this card.
Confirmed, corrected, unreadable and not-a-plate plates all count as labels. They teach Plate Review which reads it can trust.
Audit trail and reopening
Every step is recorded in an audit trail that no one can edit or delete. This includes each image being queued and read, each auto-accept (and whether the reader alone or the reader and the AI together accepted it, with both readings), each plate sent to review because the AI differed or didn't answer, each confirmation or correction (with the old and new text), each spot-check result, and each settings change with its before and after values. Every entry shows who or what made the change and when.
- A read's detail page shows that read's full history.
- Audit searches the whole trail by date, event, person, batch or image, and exports it to CSV.
- Reopening. To change a read that's already been decided, open it and choose Reopen. You must give a reason. The read goes back to review, and the reopening and its reason are recorded in the audit trail.
Settings
Changing a setting creates a new settings version. The version history shows every version, who made it, why, and the accuracy measured at the time.
- Auto-accept switches. Accept when the reader and the AI agree; let the AI file what isn't a plate; accept plates the reader alone is sure of. All start on.
- Target accuracy. How often plates the reader accepts alone must be right (default 99.5%). A higher target sends more reads to review.
- AI checks a minute. How many plates the AI reads each minute (see above).
- Spot-check rates. The share of automatic decisions also sent to review: 5% for the AI's decisions, 1% for reader alone, by default.
- Storage sources. The connected storage folders Plate Review watches (see Connected storage).
Exports
- Reads (CSV). Every read, or the reads of one batch or date range, with its text, status, scores and who reviewed it.
- Training set. The plate crops your team has labeled, with their confirmed text. Use it to measure or improve plate-reading models of your own.