The Label Image

The detector finds regions with connected components, so a label image is the result. Contours, polygons and per-region masks were all ways of describing the same thing less directly, and input/publish was a menu for choosing between descriptions. One output cannot disagree with itself.

pixel-type is Mono16, not Mono8 - a picture can hold more than 255 regions, and a label is a number rather than a brightness, so wrapping it at 255 would silently merge two regions into one.

The numbers are the blobs[].label values in output/result. That is what makes the image useful rather than decorative: from a reported area you can reach the exact pixels it was measured from.

A binary mask, if that is all you need, is labels > 0.

The image is whole frame, at the size of the input image - not the crop. Crop with crop_x/crop_y/crop_width/crop_height from output/result if you want just the searched region.

Every region output/result reports is lit here, and no other is. The picture and the numbers always describe the same regions, so a region you can see is one you can look up.

Labels are not renumbered. blobs[].label has to keep addressing the same pixels, so the numbers are the detector’s own rather than a position in the reported list — which is sorted by area, and so does not run in label order.

The Cockpit tints each region a different colour over the camera picture. That is why the label image exists rather than a plain mask: neighbouring regions come back as different colours, so a reported area can be traced to the exact pixels it was measured from. A bounding box could never show that - two overlapping boxes tell you nothing about which pixels belong to which region.