Multiple View Based 3D Object Classification Using Ensemble Learning of Local Subspaces (ThBT4.3)

1
Multiple View Based 3D Object Classification
Using Ensemble Learning of Local Subspaces
(ThBT4.3)

• Jianing Wu, Kazuhiro Fukui
• lacarte_at_cvlab.cs.tsukuba.ac.jp,
• kfukui_at_cs.tsukuba.ac.jp
• Graduate school of Systems and Information
  Engineering
• University of Tsukuba (Japan)

2
Abstract

• We proposed a statistical method for object
  classification based on multi-view.
• The proposed method is an extension of MSM.
• Problems of previous works (MSM, KMSM) has been
  solved.
• We evaluated the classification performance of
  the proposed method and previous works.

3
Table of Contents

• Backgrounds
• Multi-view classification and problems of
  existing methods for nonlinear distribution
• The proposed method
• Approximation by local subspaces and ensemble
  learning
• Experimental results
• Performance comparison using multi-view images of
  objects
• Summary

4
Multi-view Based Object Classification

• Multiple view is more beneficial for object
  classification than single.
• View based approach does not need 3D model for
  classification.
• Subspace related methods has been proposed for
  multi-view frame work.

5
Existing view based approaches

• Mutual Subspace Method (MSM)
• O.Yamaguchi, K.Fukui, K.Maeda Face recognition
  using temporal image sequence. Proc.IEEE 3rd
  International Conference on Automatic Face and
  Gesture Recognition, pp.318-323,1998.

• Pro
• Low computation cost
• Con
• Cannot handle nonlinear distribution

6
Multi-view Based Object Classification

• Feature vectors from multi-view input are likely
  to have nonlinear distribution.

• Subspace approximation therefore is not accurate.

7
Existing view based approaches

• Kernel Mutual Subspace Method (KMSM)
• H.Sakano, N.Mukawa Kernel mutual subspace
  method for robust facial image recognition. Proc.
  4th International Conference on Knowledge-Based
  Intelligent Engineering Systems and Allied
  Technologies, Vol.1, pp.245-248, 2000.
• Project feature vectors to a high dimensional
  space to reduce their nonlinearity.
• Pros
• Can handle nonlinear distribution.
• Cons
• Consume more computation time, and this increases
  at order N2 as learning data increases.
• Some critical parameters need to be optimized

8
Motivation of the Proposed Method

• Approximate nonlinear distribution.
• Divide the original distribution, and approximate
  each subset.
• Achieve comparable classification performance
  with KMSM, using less computation.
• Perform classification without kernel framework.

9
Table of Contents

• Backgrounds
• Multi-view classification and problems of
  existing methods for nonlinear distribution
• The proposed method
• Approximation by local subspaces and ensemble
  learning
• Experimental results
• Performance comparison using multi-view images of
  objects
• Summary

10
The Proposed Method

• Divide the nonlinear distribution into several
  subsets based on Euclidean distance.
• Nonlinearity of each subset is weaker.
• We approximate each subset with local subspace.

11
Ensemble Classification

• Division number and dimension of each local
  subspace are parameters affect classification
  performance.
• We construct local subspaces under multiple
  combination.
• We assume each case as weak classifier and apply
  ensemble learning.

12
The Proposed Method with Weight

• Each local subspace carries a weight coefficient
  based on its classification performance.
• This coefficient weights the canonical angle.
• The coefficients are simply the accuracy rate of
  each local subspace in preliminary experiments.

13
The Proposed Method
14
Table of Contents

• Backgrounds
• Multi-view classification and problems of
  existing methods for nonlinear distribution
• The proposed method
• Approximation by local subspaces and ensemble
  learning
• Experimental results
• Performance comparison using multi-view images of
  objects
• Summary

15
Classification Experiment

• Compare the classification performance and
  computation cost of MSM, KMSM and the proposed
  method.
• Use multi-view image data set The ETH-80 Image
  Set
• B.Leibe, B.Schiele Analyzing appearance and
  contour based methods for object categorization.
  Proc. CVPR'03, Vol.2, pp.409-415, 2003

16
Classification Experiment

• The dataset contains 8 classes, 10 objects for
  each class.

• 41 points of view for each object
• Feature vector is the resized image (16x16)

17
Classification Experiment

• Use 164 images to learn for each class. (generate
  dictionary)
• Evaluation input is images from an unknown
  object. (10 images/points of view)
• By exchanging learning data and evaluation data
  (leave-one-out), we repeated experiment for 1640
  times.

18
Classification Experiment

• Classification performance of each method

Method Accuracy Rate() Separability EER()
MSM 69.5 0.34 20
KMSM 87.2 0.41 15
Proposed method 86.5 0.44 14
Proposed method with weight 94.7 0.55 9
19
Classification Experiment

• The proposed method improved classification
  performance from MSM
• The proposed method achieved comparable
  performance with KMSM
• Classification performance of the proposed method
  was improved by introducing weight to each local
  subspace

20
Classification Experiment

• Computation cost of each method

Method Calculation Time (seconds per input) Order
MSM 0.1 O(n)
Proposed method 0.4 O(n)
Proposed method with weight 0.4 O(n)
KMSM 3.1 O(n2)
21
Classification Experiment

• The proposed method consumes less computation
  time compared with KMSM
• The computation time of the proposed method
  increases in order N as learning data increases

22
Classification Experiment

• Classification performance of weak classifiers

Proposed method 86.5 0.44 14
23
Table of Contents

• Backgrounds
• Multi-view classification and problems of
  existing methods for nonlinear distribution
• The proposed method
• Approximation by local subspaces and ensemble
  learning
• Experimental results
• Performance comparison using multi-view images of
  objects
• Summary

24
Summary

• We proposed a method to achieve comparable
  performance with KMSM by less calculation.
• Classification performance is further improved by
  introducing weight to each local subspace
• The advantages of the proposed method is shown
  with classification experiment of objects.

25
Thank You

• Multiple View Based 3D Object Classification
  Using Ensemble Learning of Local Subspaces
  (ThBT4.3)
• Jianing Wu, Kazuhiro Fukui