2

I need help determining if what I'm doing is sufficient. I have extracted the "curvature" from two human faces (essentially the curve that goes from one ear, down the face to the chin, and back up to the other ear). Now I want to compute a curve similarity with another human face.

To try to make this invariant to the possible different curve sizes (since the images these faces are taken from can be taken from varying distances from the person) I've normalized both the X and Y dimensions by subtracting the mean and dividing by the standard deviation (I'm not sure this is correct).

Finally, I'm using Dynamic Time Warping to compare these curves, this is the kind of result I get:

DTW on Face Curves

Basically I'm asking whether this approach sounds reasonable.

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467
user2974849
  • 193
  • 1
  • 8
  • Not sure about DTW, but I have seen people using [Procrustes analysis](http://en.wikipedia.org/wiki/Procrustes_analysis) for related morphometric analyses. – chl Nov 09 '13 at 21:39
  • If you're interested in comparing shapes (say as characterized by curvature in some sense), then wouldn't DTW change the very things you're interested in measuring? – Glen_b Nov 10 '13 at 02:30
  • I think it does. I've since taken chl's suggestion and am doing Procrustes analysis instead. Thanks for your help. – user2974849 Nov 10 '13 at 03:50

0 Answers0