1

There are many CNNs for object classification and shapes drawing (e.g., square, heart, curved line etc) classification is even simpler.

However, I haven't found yet a robust method for shape quality classification. I.e., given a drawing of a circle, I want to know how similar it is to a perfect circle. for example, duplicate vs original circle

Now, there is a SSIM index but it was designed to distinguish between photos and not drawings. This is why the position of the shape in the image (within a few pixels) can cause use SSIM differences. I attach some result so you can see what I mean diffrent SSIM for diffrent circles

The method Im using right now is to hand-label the shapes and perform a supervised network, but I wonder maybe there is a better way (maybe siamese networks?

Roi Yozevitch
  • 197
  • 3
  • 13
  • There are lots of simple image analysis methods to quantify similarity to a circle, you don’t need a classifier for this. Look for example into the circularity or roundess measures. – Cris Luengo Aug 02 '22 at 15:41
  • Thanks but it doesn't help me. the circle was just an example. one can look at the triangles to see the point. and when we have more complex shapes it doesn't work. – Roi Yozevitch Aug 03 '22 at 10:29

0 Answers0