1

i was looking into face recognition and i came across this method to face transform using similarity transform to align faces detected by mtcnn for the megaface dataset, in that the source matrix being used was

src = np.array([                                                                                                                                                                     
 [30.2946, 51.6963],                                                                                                                                                                
 [65.5318, 51.5014],                                                                                                                                                                
 [48.0252, 71.7366],                                                                                                                                                                
 [33.5493, 92.3655],                                                                                                                                                                
 [62.7299, 92.2041] ], dtype=np.float32 )

can anyone tell me from where do we derive this matrix? and/or why this is used?

reference: https://zhuanlan.zhihu.com/p/29515986

Imtinan Azhar
  • 1,725
  • 10
  • 26
  • Isn't there any English source to read? – Eypros Sep 05 '18 at 06:59
  • i just used chromes google translate that converts it to english, i can refer you to a code file that helps, https://github.com/deepinsight/insightface/blob/master/src/align/align_facescrub.py – Imtinan Azhar Sep 05 '18 at 07:08
  • i have seen multiple instances use this src matrix, i know that these are landmarks, but why do they get used – Imtinan Azhar Sep 05 '18 at 07:09

1 Answers1

0

I figured it out, basically, the points are the landmark points of a face that resides in a 112,96 image, the nose is centered in the image and thus transforming landmark points of another face onto these points will center align a face just like these template landmark points

insanely_sin
  • 986
  • 1
  • 14
  • 22
Imtinan Azhar
  • 1,725
  • 10
  • 26
  • How are you concluding nose is always at the center of the image after transformation? – insanely_sin May 26 '19 at 03:07
  • This is because we have a predicted set of points, the points are given to us by MTCNN, we transform and warp these points onto the above-mentioned src points on a 112x96 image, this centers the face of the image, by mapping the nose of the original face to the nose(center) of the target image – Imtinan Azhar May 26 '19 at 07:27