1

I have the following task: recognize a set of simple hand-drawn shapes on a sheet of paper from a still image (not a video stream), so they might not be exactly identically pixelwise.

Those shapes are basically symbols for doors, windows, etc. in a floor plan (see attached image), so they might be slightly scaled or rotated (90° steps). There are about 5 different ones.

So far I came across SIFT (and its OpenCV-variants SURF and ORB) as well as a cascaded classifier to recognize haar-like features.

For SIFT there seem to be too little key points in such a shape whereas I did not manage to get the haar-trained cascaded classifier to work. Also, a cascaded classifier seems a bit heavy for recognizing such simple shapes, no?

Does anyone of you have any good hints or alternative approaches? Or maybe you even have a snippet of code lying around which I can use?

Symbol for a door

Paul Latzelsperger
  • 711
  • 1
  • 6
  • 8

1 Answers1

1

I think histograms of gradients (HOG) should work great for such elements.

Andrey Smorodov
  • 10,649
  • 2
  • 35
  • 42