Outline • A. M. Martinez and A. C. Kak, “PCA versus LDA,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 23, no. 2, pp. 228-233, 2001. The Goal • The main goal of the paper is to examine where LDA is always superior to PCA – This is a natural tendency as LDA is designed for discrimination while PCA is not May 29, 2016 Computer Vision 2 Comparison of PCA and FDA May 29, 2016 Computer Vision 3 Illustration • LDA vs. PCA LDA PCA DLDA DPCA PCA May 29, 2016 LDA Computer Vision 4 LDA PCA DLDA DPCA PCA LDA PS: The decision thresholds, marked DPCA and DLDA, is yielded by the nearest-neighbor approach May 29, 2016 Computer Vision 5 Localization and Morphing of face image • The paper compares PCA and LDA with regard to only the identification of faces. – Localization step: • A face was first manually localized by marking the left, the right, the top, and the bottom limits of the face, as well as the left and the right eyes and the nose – Morphing step: • After localization, faces are morphed so as to fit a grid of size 85 by 60 May 29, 2016 Computer Vision 6 Localization and Morphing of face image Localization Morphing 60 85 May 29, 2016 Computer Vision 7 Segmentation • Each image is segmented by means of an oval-shaped mask centered at the middle of the morphed image rectangle – There are t pixels in the oval-shaped segment t-dimensional vector xi – ( N sample images X = { x1 , . . . , xN } ) May 29, 2016 Computer Vision 8 Experimental Results • AR-face database – This database consist of 126 classes (people), and there are 26 different images per class. – For each class, these images were recorded in two different sessions separated by two weeks, each session consists of 13 images. – 50 different classes (25 males, 25 females) were randomly select from database. – As state earlier, images were morphed to the final 85x60 pixel array, segmented using an ovalshaped mask. May 29, 2016 Computer Vision 9 One Subject May 29, 2016 Computer Vision 10 Small Training Data Sets • For each classes, we use only 7 image (a ~ g) – 2 images for training and 5 images for testing – There are in total 21 different ways of separating the data for each class • They are labeled as Test#1 , Test#2 , . . . , Test#21 May 29, 2016 Computer Vision 11 Small Training Data Sets May 29, 2016 Computer Vision 12 Small Training Data Sets May 29, 2016 Computer Vision 13 Small Training Data Sets May 29, 2016 Computer Vision 14 Small Training Data Sets May 29, 2016 Computer Vision 15 Small Training Data Sets May 29, 2016 Computer Vision 16 Small Training Data Sets May 29, 2016 Computer Vision 17 Small Training Data Sets May 29, 2016 Computer Vision 18 Using a Larger Training Set May 29, 2016 Computer Vision 19 Using a Larger Training Set May 29, 2016 Computer Vision 20 Experimental Results (cont.) • For large training sample: – 13 images of first session for training – 13 images of second session for testing May 29, 2016 Computer Vision 21 Conclusion • PCA and LDA have been demonstrated to be useful for many application such as face recognition. • One might think that LDA should always outperform PCA (since it deals directly with class discrimination) • The experiments we report here suggest otherwise. • When PCA outperforms LDA, the number of training samples per class is small May 29, 2016 Computer Vision 22