Use the nearest neighbor recognition algorithm to perform face recognition on the dataset of the faces.zip file that you will find on Eleum. In particular, ORL_32x32.mat contains 10 32x32 different images for 40 people. For some subjects, the images were taken at different times, varying the lighting, facial expressions (open / closed eyes, smiling / not smiling) and facial details (glasses / no glasses). Moreover, you will find three more folders (3Train, 5Train, 7Train) which account for the number of training images per person (to be used in your project) and are actually the corresponding indices for each one of the 400 images, classifying them as training or testing samples.
Before applying the method, scale all image intensities so that they range from 0 to 1
- Apply eigen-analysis as discussed in the lecture (and also the book of Szeliski) and plot out the first K principal components (use the reshape function).
- Project the training data on the new space and derive descriptors w tr,i for each training image i. Show how a trained image can be reconstructed using eigenvectors and w tr,i . This reconstruction will help you chose the right K for your experiments.
- Project the test data and derive descriptors w te,i .
- Implement a NN technique for face recognition using your descriptors. Discuss the results and the accuracy you obtain (fraction of test images correctly classified). Comment on your results for a varying number of K (indicatively, 10, 20, 30) and a varying number of training samples per person. Create plots for each classification rate.
$ python3.6 main.py