Epitome Ji Soo Yi and Woo Young Kim Instructor: Prof. James Rehg April 27, Spring 2004, CS7636 Computational Perception
CONTENTS Introduction Epitomic Image Experiment Results & Conclusion Future direction Edited by Woo Young and Ji Soo
Introduction(1) Image representative model Feature-based Geometric approach Template-based Standard Euclidian error norms Eigen spaces Color histogram-based Edited by Woo Young and Ji Soo
Introduction(2) Epitomic image analysis What is Epitome? The miniature, condensed version of image. Still consists of most constitutive elements. Use a probabilistic measure of similarities. Shape epitome and appearance epitome. Edited by Woo Young and Ji Soo
Introduction(3) Epitomic image analysis Graphical model of epitomic analysis Edited by Woo Young and Ji Soo EsEs M S1S1 S2S2 I EmEm appearance epitome shape epitome I=M*S 1 +(1-M)*S 2 + noise
Introduction(4) Epitomic image analysis Probabilistic framework Edited by Woo Young and Ji Soo epitome e = ( , ) M,N Patch Z k = {z i,k }, z i,k = x i Input image X Patch Z n Me, Ne TkTk TnTn
Introduction(5) Epitomic image analysis EM algorithm to extract an epitomic image Edited by Woo Young and Ji Soo E step: M step:
Epitomic Image (1) Edited by Woo Young and Ji Soo Original image Epitomic image
Epitomic Image (2) Edited by Woo Young and Ji Soo Input image Epitomic image
Experiment (1) Edited by Woo Young and Ji Soo Epitomic Modeling Face Detection Comparison with PCA Analysis
Experiment (2) Edited by Woo Young and Ji Soo Epitomic Modeling Training data – a set of face images Each image : 100 by 75 Epitomic image: 32 by 32 Epitomic image
Experiment (3) Edited by Woo Young and Ji Soo Epitomic Modeling Training data – a synthetic image by tiling face images 100 by 75 pixels for each image 1000 by 375 pixels for total 75 by 75 pixels
Experiment (4) Edited by Woo Young and Ji Soo Face Detection Histogram and clustering
Experiment(5) Edited by Woo Young and Ji Soo Face Detection Patch matching – face image High log likelihood – good matchLow log likelihood - poor match
Experiment(6) Edited by Woo Young and Ji Soo Face Detection Patch matching – non face image Low log likelihood – good matchHigh log likelihood - poor match
Experiment(7) Edited by Woo Young and Ji Soo Comparison with PCA analysis – PCA Rigid data Non-Rigid data
Experiment(8) Edited by Woo Young and Ji Soo Comparison with PCA analysis – Epitome Rigid data Non-Rigid data
Results & Conclusion Edited by Woo Young and Ji Soo Epitomic image modeling Parameter settings Comparison with PCA Analysis Statistics
Future direction Edited by Woo Young and Ji Soo Computational time saving Shape epitome Other applications