That's what those 3 corners are for.
After a convolution with such a square as template, there are 3 local maxima, of known relative distances to each other, that give the bounding box.
this post was submitted on 28 Mar 2025
7 points (88.9% liked)
Computer Vision
50 readers
1 users here now
For the computer vision community
founded 3 weeks ago
MODERATORS
Is that scale invariant? Or would they need to do it multiple times at different scales if they can't rely on a consistent scale?
The latter: it's not scale invariant
Wouldn’t it be easier to generate a new one?