|
Assistent: Daniela Hall
Face recognition with principle components analysis Please download the test images, a color map for correct display and a example program: pca.tar.gz For exercise No.5, please download the images here: coil.tar.gz
Demonstration on interest point detectors
For additional reading: Perception of local image appearance Thesis_chap3.ps.gz Solutions of the exercise: PCA for face recognition
Local PCA
Mean:The mean image Phi of 18 face images
Diff:The difference images I-Phi
PCA:The eigenvectors or principal components of 18 face images ordered by decreasing eigen values
Recons1:Reconstruction of a face image from the training set using 15, 16, or 17 eigen vectors. Observation: The reconstruction with 17 eigen vectors is perfect. The reconstruction using fewer eigen vectors shows some blurr.
Recons2:Reconstruction of a face that is not in the training set. Observation: The reconstruction adds glasses to the image. The reason for this is that all training images showed faces with glasses.
Recons3:Reconstruction of a image that is not a face. Observation: We observe a large reconstruction error. PCA can only reconstruct the projection of the image into face space. When the image is not a face, the difference from face space (DFFS) is large.
LocalPCA:Training images for local PCA
LocalMean:The mean of all local subimages of size 9x9. Observation: the mean is uniformly grey, because the average of a very large number of subimages converges to a uniformly grey image.
LocalEV:The eigen vectors of local PCA orderd by decreasing eigen value. Observation: The eigen vectors converge towards derivatives.  |
 |