0

Ive been exploring dlib's face detector over its python API. On most images in my data set it seems to perform slightly better than cv2 on most images so I kept playing around with it on multiple faces in picture scenarios.

Going through dlib's python examples it seems like it would be possible to train these images but I am wondering if anyone has a suggestion how to make sure that the two faces on the far left and right are detected out of the box?

This is he image that I am having trouble finding all 6 faces on (https://images2.onionstatic.com/onionstudios/6215/original/600.jpg)

Dennis
  • 81
  • 1
  • 9

2 Answers2

2

Dlib has a very precise face detector. But it works bad detecting not frontal (like far left) and/or occluded faces (like far right).

Seeta (https://github.com/seetaface/SeetaFaceEngine) works better with those. But it's less precise.

Also I tried retraining Dlib's face detector. And obtained much lower precise than DLIB and less recall than Seeta. So, re-training DLIB seems not perfect idea.

Alexey Antonenko
  • 2,389
  • 1
  • 18
  • 18
  • were you able to find a different solution that would cover the same issue? – Dennis Dec 29 '17 at 07:07
  • Using Yolo tiny object detection model I made transfer learning and obtained face detection model which has optimal performance (good precision and recall). But it uses GPU to provide ~20 msec / image processing time. Other approaches I'm trying now: 1) use two detectors (like Seeta + DLIB), 2) try other detectors (like MTCNN, Pico), 3) use filtering to reject detections that may be wrong (like blurryness estimation). – Alexey Antonenko Jan 02 '18 at 15:42
  • oh awesome, I was just reading about YOLO the other day. Ill give it a try thanks! – Dennis Jan 04 '18 at 04:49
  • But to be honest, I had success with that just one. Later I tried a few more times to repeat that but for some reason with no success. – Alexey Antonenko Jan 04 '18 at 10:17
0

In my experience, Dlib does not do very well out of the box with obscured and profile faces out of the box. I would recommend training Dlib with more data of this kind.

Gautham Velchuru
  • 65
  • 6
  • Thats interesting, looking around the web a couple blog posts were raving about how good dlib was specifically for these non-frontal pictures. Thanks – Dennis Jul 03 '17 at 22:52