We are using YOLO11 segmentation to detect inner and outer contours of a pipe cross-section. The goal is to measure inner diameter, outer diameter and wall thickness in millimeters. The challenge is that segmentation masks are not perfectly aligned with the true contour, causing measurement errors. Is segmentation-based metrology sufficient, or should we combine segmentation with edge detection, ellipse fitting, camera calibration and subpixel contour extraction?
I don’t know anything about segmentation-based metrology, so I can’t really help with that. If I were trying to measure the pipe inner/outer/wall dimensions, I’d probably be thinking about things like:
- Using a telecentric lens (no perspective distortion, so scale changes with distance). But telecentric lenses are expensive, particularly large ones. (how big is that pipe?)
- controlling your lighting / background to improve detection of edges / boundaries.
- Image processing → edge detection → circle / ellipse fitting. (But someone is probably going to say that you can just throw AI at this with a decent training set and be done in time for dinner. And they might be right.)
- What about getting rid of the cameras altogether? If you had distance sensors (ultrasonic? laser?) attached to a fixture that you could put inside / around the pipe and then measure the distance to the inner/outer wall in a few spots, you could get all of your dimensions.