150 likes | 236 Views
Learning Visual Similarity Measures for Comparing Never Seen Objects. By: Eric Nowark , Frederic Juric Presented by: Khoa Tran. Outline. 1.) Purpose 2.) Methodology 3.) Results. Purpose. Object Recognition. Train Images. Test Images. Methodology Preview.
E N D
Learning Visual Similarity Measures for Comparing Never Seen Objects By: Eric Nowark, Frederic Juric Presented by: Khoa Tran
Outline • 1.) Purpose • 2.) Methodology • 3.) Results
Purpose Object Recognition Train Images Test Images
Methodology Preview A.) Corresponding patch pair B.) Quantizing patch pair C.) Patch pair similarity measure
Object Recognition • 1.) Images • 2.) Feature Extraction • 3.) Model Database • 4.) Matching • a.) Hypothesis Generation • b.) Hypothesis Verification Images Features Extraction Model Database Matching Hypothesis Generation Hypothesis Verification
Images • Total: - 225 images, - 21 different objects • Training Data Set - 1185 positive image pairs - 7330 negative image pairs - 14 different objects • Testing Data Set - 1044 positive image pairs - 6337 negative image pairs - 7 different objects
Feature Extraction • Patches • Normalized Cross Correlation • SIFT Descriptors • Matrix representation
Model Database • Extremely Randomized Binary Decision Tree • SIFT Descriptors • Geometric Information • Information Gain
Hypothesis Generation – Similar Measure • Similar Measure • Support Vector Machine
Hypothesis Generation Ferencz and Malik Faces in the News Dataset Dataset
C.) Hypothesis Verification • Sammon mapping for toy cars
Results 1.) Toy Cars 2.) Ferencz 3.) Faces 4.) Coil 100
Reference • Eric Nowak and Fredric Jurie; "Learning Visual Similarity Measures for Comparing Never Seen Objects” ;Computer Vision and Pattern Recognition 2007 (CVPR'07);