This paper demonstrates an unsupervised face annotation algorithm.
As I mentioned in [Reading] lecture 4, a major difference between PCA and LDA (I used the term FLD at that time) is that the LDA exploits supervised data set to do dimension reduction. It makes LDA perform better than PCA for classification, but at the cost of the supervision beforehand. However, this paper still can utilize LDA without labeling beforhand! The reason is that a lexicon of names can be extracted automatically from the news captains. Those names may not be accurate for all images, but are already enough for using LDA to do dimension reduction (or we can say obtaining distinctive property).
They didn't restrict the dimension reduction method to LDA, Kernel-PCA is also recommended. However, their database is too large to do Kernel-PCA, so they use Nystrom Approxiamtion to make this job easier.
After doing dimension reduction, clustering is done by a modified K-Means method. This can be interpreted as a face labeling process. Since the lexicon of names obtained at the beginning is noisy, they futher prunes and merges the clusters to make the result better.
The paper gives a "face recognition" algorithm which is much different to the previous works. It utilizes the captain information to automatically generate the supervised dataset (and also the test dataset). However, I think their system can only recognize faces that appear several times.
Reference:
T. L. Berg and A. C. Berg and J. Edwards and M. Maire and R. White and Y. W. Teh and E. Learned Miller and D. A. Forsyth, "Names and faces in the news," IEEE Computer Vision and Pattern Recognition or CVPR, p. II: 848-854, 2004.
As I mentioned in [Reading] lecture 4, a major difference between PCA and LDA (I used the term FLD at that time) is that the LDA exploits supervised data set to do dimension reduction. It makes LDA perform better than PCA for classification, but at the cost of the supervision beforehand. However, this paper still can utilize LDA without labeling beforhand! The reason is that a lexicon of names can be extracted automatically from the news captains. Those names may not be accurate for all images, but are already enough for using LDA to do dimension reduction (or we can say obtaining distinctive property).
They didn't restrict the dimension reduction method to LDA, Kernel-PCA is also recommended. However, their database is too large to do Kernel-PCA, so they use Nystrom Approxiamtion to make this job easier.
After doing dimension reduction, clustering is done by a modified K-Means method. This can be interpreted as a face labeling process. Since the lexicon of names obtained at the beginning is noisy, they futher prunes and merges the clusters to make the result better.
The paper gives a "face recognition" algorithm which is much different to the previous works. It utilizes the captain information to automatically generate the supervised dataset (and also the test dataset). However, I think their system can only recognize faces that appear several times.
Reference:
T. L. Berg and A. C. Berg and J. Edwards and M. Maire and R. White and Y. W. Teh and E. Learned Miller and D. A. Forsyth, "Names and faces in the news," IEEE Computer Vision and Pattern Recognition or CVPR, p. II: 848-854, 2004.
沒有留言:
張貼留言