40
Kernel Methods for Implicit Su rface Modeling B. Scholkopf, J. Giesen, S. Spalinger r92922120 黃黃黃 r93922020 黃黃黃 r93922038 黃黃黃 r93944009 黃黃黃

Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Embed Size (px)

DESCRIPTION

Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger. r92922120 黃邦洪 r93922020 楊惠菁 r93922038 柯政宏 r93944009 劉弘偉. Outline. Introduction Single-Class SVMs Slab SVMs Epsilon-SVR Implement Result. Introduction. Step 1: - PowerPoint PPT Presentation

Citation preview

Page 1: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Kernel Methods for Implicit Surface Modeling

B. Scholkopf, J. Giesen, S. Spalinger

r92922120 黃邦洪r93922020 楊惠菁r93922038 柯政宏r93944009 劉弘偉

Page 2: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Outline

Introduction Single-Class SVMs Slab SVMs Epsilon-SVR Implement Result

Page 3: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Introduction

Step 1: Given a finite sample points belong to some

hypersurface in Euclidean space R3

The sample can be very large, noisy or sparse

Page 4: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Introduction

Step 2: Use SVM to find an implicit surface

f is the decision function in SVM

The zero set of f is a surface that approximates

the sample points

Single-Class SVM v.s. Slab SVM

3:f

)0(1f

Page 5: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Outline

Introduction Single-Class SVMs Slab SVMs Epsilon-SVR Implement Result

Page 6: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Single-Class SVM

Primary Object function

Concept

Page 7: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Single-Class SVM

Dual Problem

Decision function

Page 8: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Single-Class SVM

SVM TOY example

Page 9: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Outline

Introduction Single-Class SVMs Slab SVMs Epsilon-SVR Implement Result

Page 10: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Slab SVM

Primary object function

Concept

Page 11: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Slab SVM

Lagrangian dual optimization problem

Decision function

Page 12: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Slab SVM

SVM TOY example

Page 13: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Outline

Introduction Single-Class SVMs Slab SVMs Epsilon-SVR Implement Result

Page 14: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Epsilon-SVR

Primary object function

Concept

Page 15: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Epsilon-SVR

Lagrangian dual optimization problem

The approximate function

Page 16: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Outline

Introduction Single-Class SVMs Slab SVMs Epsilon-SVR Implement Result

Page 17: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Implement

Single-Class SVM Use LIBSVM directly

Slab-SVM Modify epsilon-SVR in LIBSVM to slab-SVM The result is parameter-sensitive Modification:

decision function

The result is good but we don’t really know how to explain

Page 18: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Outline

Introduction Single-Class SVMs Slab SVMs Epsilon-SVR Implement Result

Page 19: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(1) - One-Class SVM

Bunny Original: 35947 One-Class: 2810

Page 20: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(1) - Slab SVM

Bunny Left: 8307。 Right: 1013

Page 21: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(1) - Slab SVM

Bunny Predict accuracy: 97.74%

Page 22: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(2) - One-Class SVM

Cactus Original: 3337 One-Class: 1617

Page 23: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(2) - Slab SVM

Cactus Left: 996 Right: 817

Page 24: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(2) - Slab SVM

Cactus Predict accuracy: 96.9434%

Page 25: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(3) - One-Class SVM

Dinosaur Original: 56194 One-Class: 44963

Page 26: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(3) - Slab SVM

Dinosaur Left: 13005。 Right: 1018

Page 27: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(3) - Slab SVM

Dinosaur Predict accuracy: 97.43%

Page 28: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(4) - One-Class SVM

Knot Original: 10000 One-Class: 8807

Page 29: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(4) - Slab SVM

Knot Left: 1176。 Right: 1036

Page 30: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(4) - Slab SVM

Knot Predict accuracy: 96.9%

Page 31: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(5) - One-Class SVM

Screwdriver Original: 27152 One-Class:21728

Page 32: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(5) - Slab SVM

Screwdriver Left: 6574 Right: 884

Page 33: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(5) - Slab SVM

Screwdriver Predict accuracy: 97.1752%

Page 34: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(6) - One-Class SVM

Rockerarm Original: 40177 One-Class: 4868

Page 35: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(6) - Slab SVM

Rockerarm Left: 2090 Right: 1029

Page 36: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(6) - Slab SVM

Rocker arm Predict accuracy: 96.18%

Page 37: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(7) - One-Class SVM

Hole Original: 4000 One-Class: 3224

Page 38: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(7) - Slab SVM

Hole Left: 648 Right:1039

Page 39: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Result(7) - Slab SVM

Hole Predict accuracy: 93.475%

Page 40: Kernel Methods for Implicit Surface Modeling B. Scholkopf, J. Giesen, S. Spalinger

Q&A

Thank You