AI Detection
A computer-vision model looks at submitted photos and draws a box around the problem. It is a suggestion, not a verdict: a human moderator confirms it before it is used to teach the model anything.
Thousands of photos cannot be looked at by hand, and a photo is only useful if you know where in it the problem is.Basis uses an object-detection model to find the problem in the frame, so photos can be sorted, checked and learned from at scale.
Step by step.
A photo comes in
When a signed-in citizen files a report with a photo, the photo is stored and queued for analysis. Filing the report never waits on the analysis.
The model looks for known problems
A YOLOv8 object-detection model scans the image and returns boxes with a confidence score for each thing it recognises.
Boxes are matched to the category
The citizen's chosen category is compared with what the model saw, so a pothole report with a pothole box is easy to tell from one where the model found nothing.
A moderator takes it from there
Every result, including no detection at all, goes to the human review queue.
The details.
Eleven things it looks for
Pothole, garbage, dumpster, road crack, water logging, manhole, broken street light, sidewalk damage, graffiti, billboard damage and soil erosion.
Boxes and confidence
Each detection has a position and a confidence score, so uncertain results are visible rather than hidden.
Honest about its limits
Traffic violations and the catch-all category are not things the model can see yet. Those reports rely on human review.
One shared model
The same model serves the citizen app and the internal street-survey tooling, and improves through review.
- Model
- YOLOv8 object detection
- Recognises
- 11 classes of urban issue
- Output
- Bounding boxes with confidence scores
- Runs on
- Signed-in reports today
Human Review & Continuous Learning
Moderators confirm or correct the AI. Corrections teach the model.
Read more →Anonymous Reporting
Photo, pin, category. No account, name or phone number.
Read more →Rover Network
Autonomous ground rovers as a physical verification layer.
Read more →See what it finds
Report a problem and see it through to review.