Title: 2. Classification of normal vs abnormal
1LEARNING THE BEST SIMILARITY MEASURES FOR 3D
OBJECT RETRIEVAL AND CLASSIFICATION Department of
Computer Science and Engineering, University of
Washington Indriyati Atmosukarto, Prof. Linda
Shapiro Indria, shapiro_at_cs.washington.edu
MOTIVATING APPLICATIONS
3. Shape-Based Similarity Retrieval of Human
Faces
1. Classification and retrieval of general 3D
objects
2. Classification of normal vs abnormal
mouse skulls with cleft lip/palate genes
Abnormal skull with two gene KO
Normal skull
Abnormal skull with one gene KO
Horse vs Tiger
Class Q1000 vs Q1001
OBJECTIVE
APPROACH
3D OBJECT REPRESENTATION
Mid level Feature aggregator For scale
s Histogram Spin image Clustering
Low level Feature extraction Gaussian
curvature Mean curvature Shape index First order
derivative Curvedness
3D Meshes class Q1000
Every point pi has value vi
3D Meshes class Q1001
Eigenvectors Eigenvalues
Coefficient clusters
Every point has coeffs
Coefficient Clustering (K-means)
PCA
Calculate coefficient For all points
Every point pi has vector fi
Segmented object
NEXT QUESTION which regions are useful for
discrimination?
This work is supported by NSF Grant No.
DBI-0543631