For this configuration, I used the frontal (SET 1) images to make the average image. After building the face space from images in the database using SVD, the 20 best basis vectors returned by SVD is used for recognition. Model for each person in the databse is created by averaging the basis space coefficients for each of the 01,03,and 04 images. To match new images , the input image is first converted to a set of basis coefficients. In order to identify non-face input images, these basis cofficients are projected to face space and thus recreating a basis image. I manually picked the threshold for non-face images. The comparison was done by taking the SSSD between the original image and the reprojection. If the resulting euclidean distance is within threshold, then compare the N basis coefficients with those of the models and return the best match.

The pre-processing of images for this configuration involves cropping 35 pixels at the bottom in the vertical direction. Otherwise, it is similar to configuration 1.