2

I have extracted faces from a video and I clustered them in big groups (each group contains faces from the same person, I did this using change of background detection). Now I want to cluster those groups into a smaller number of groups and to have, at the end, one cluster per person appearing in the video.

At the moment I managed to cluster those groups using histograms of colors and K-means. The problem is that when skin and hair color are too close for two different people, the clusters are bad. Which other features can I use for this? I tried DAISY descriptors but the results were not interesting. Also I would like to make those clusters without having to specify the number of clusters I want to have at the end.

Miki
  • 40,887
  • 13
  • 123
  • 202
N. Ruchers
  • 147
  • 1
  • 11
  • Have a look [here](http://docs.opencv.org/2.4/modules/contrib/doc/facerec/facerec_tutorial.html) – Miki Apr 06 '16 at 19:32
  • I may be mistaken but I think face recognition doesn't fit with what I want to do. I don't have any training set, I want to do this in an unsupervised way if it is possible. – N. Ruchers Apr 06 '16 at 19:36

1 Answers1

0

To make clusters without having to specify the number of clusters you want to have at the end, you can use MST based clustering : https://sites.google.com/site/dataclusteringalgorithms/mst-based-clustering-algorithm

To have one cluster per person appearing in the video, you need to build better similarity function between faces. State-of-the-art face recognition systems uses Convolutional Neural Network approach. You can read more about it in Facebook article : https://www.cs.toronto.edu/~ranzato/publications/taigman_cvpr14.pdf.

taarraas
  • 1,483
  • 9
  • 18