Smart Detections are possible defects the AI spotted in your images, such as cracks or corrosion, depending on the models run on your project. The AI is a first pass over thousands of images, and your review is what turns it into an inspection record.
They are suggestions, not observations. Nothing enters your inspection record until you review a detection and convert it. See Accepting, dismissing or converting a detection.
Switch to AI Review with the mode picker in the header. The detections are listed in its Smart Detections panel. If the mode is not there, your plan does not include AI processing, or none was run on this project.
By class, not as one long list. Each row shows the class, how many detections it holds, and how much of it you have reviewed. You work through one kind of defect at a time, taking all the cracks and then all the corrosion, rather than switching context on every item.
Each detection carries a confidence percentage: how sure the model is that it found what it says.
The meter is monochrome. A green-amber-red scale would imply a threshold above which you can trust the result and below which you can ignore it. No such threshold exists. A 60% detection can be a real crack, and a 95% one can be a shadow.
Treat confidence as a sorting hint, not a verdict. It tells you where to start, not what to conclude.
| Mark | Meaning |
| Hollow circle | Not yet reviewed |
| ✓ | Accepted as real |
| ✗ | Dismissed as not real |
Shape and colour both encode this, so it stays readable if you have difficulty distinguishing colours.
All classes takes you back to the overview.
Each class shows how much of it you have reviewed, and the bar completes at 100%. With several people reviewing, or a job spread over days, that is how you find where to resume.
To narrow what you see, filter by status: All, Unrated or Reviewed. Unrated is the useful one for a second sitting, because it hides everything already judged so you see only what is left. The filter is per session and resets when you leave.
The image zooms and pans like any other, so you can inspect the detected area properly before deciding.
Where a detection appears in several images, a thumbnail strip lets you switch between them, and chevrons step through. Seeing the same spot from two angles often settles an ambiguous case.
Go to detection in the menu flies the 3D camera to where that image was taken, which puts the detection in the context of the structure.
Recording a verdict does not move you to the next detection. You choose the next one yourself, so you can record a verdict and keep looking at the same image.
Select an observation in Inspect, switch to AI Review, and its details are gone. Nothing has been lost. The observation details panel exists in Inspect only, and AI Review uses that space for the detection you are reviewing: its image, confidence and verdict controls. Report mode uses the same space for its own content.
Your selection survives the switch. Switch back to Inspect and the observation is still selected, with its details exactly as you left them. So you can look at something in Inspect, switch over to see whether the AI flagged the same spot, then come back with your place intact.
| What you see | What it means |
| No AI Review in the mode picker | Your plan does not include AI processing, or none was run on this project. |
| An observation's details vanished when you switched mode | The details panel is in Inspect only. Switch back and it reappears. |
| Recording a verdict left you on the same detection | Select the next detection yourself. |