23
Andreas Opelt (Graz University of Technology and University of Leoben) 1 A. Opelt, M. Fussenegger, A. Pinz, P. Auer Weak Hypotheses and Boosting for Generic Object Detection and Recognition

A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Embed Size (px)

DESCRIPTION

A. Opelt, M. Fussenegger, A. Pinz, P. Auer. Weak Hypotheses and Boosting for Generic Object Detection and Recognition. Agenda. The Basic Idea Our Framework for generic Object Recognition  The techniques used The Learning Model  Our Model The Weak Hypotheses Finder Experiments - PowerPoint PPT Presentation

Citation preview

Page 1: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

1

A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Weak Hypotheses

and Boosting

for Generic Object Detection

and

Recognition

Page 2: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

2

Agenda

• The Basic Idea• Our Framework for generic Object Recognition • The techniques used• The Learning Model

• Our Model• The Weak Hypotheses Finder

• Experiments• Discussion / Outlook

Page 3: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

3

The Basic Idea 1/2

We want to go towards ‘real’ Generic Object Recognition!

No pre-selection of the object !

Arbitrary view of the object!

Any instance of the object category!

Any background clutter!

Object is located anywhere in the

image!

Objects shown in any arbitrary scale!

Not only for a special category of objects!

Not special images for learning!

Problems ?

Page 4: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

4

Agarwal and Roth, ECCV 2002,

Cars side database

The Basic Idea 2/2

We want to go towards ‘real’ Generic Object Recognition!We want to go towards ‘real’ Generic Object Recognition!

Oxford database; (Fergus, Perona and Zisserman, CVPR 2003)

Graz database; Bikes, Persons, Background

Page 5: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

5

The Framework

Page 6: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

6

Region Extraction 1/2

[Mikolajczyk/Schmid 2001]

Data reduction: Threshold

[Mikolajczyk/Schmid 2001]

Data reduction: Threshold

[Lowe 1999] (Diff. of Gaussian)

Data reduction: Clustering

Page 7: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

7

Region Extraction 2/2

Page 8: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

8

Region Normalization

• Homomorphic Filtering [Gonzales and Woods, C. 4.5.]

• Size Normalization

Page 9: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

9

The Framework

Page 10: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

10

Local Descriptors

Subsampled Grayvalues

Basic Moments (Dim=10)

[L. Van Gool 1996]

Dim=9

[D. Lowe 1999]

Dim=128 (3 orient. planes, 8x8px)

Page 11: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

11

The Framework

Page 12: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

12

The Learning Model 1/3

Input:

Output:

Weak Hypotheses:

Threshold, Weight

Page 13: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

13

The Learning Model 1/3

Select best Weak Hypothesis

Calculate Threshold

Page 14: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

14

The Learning Model 3/3

Page 15: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

15

Experiments 1/6

Category: Bikes some Weak Hypotheses

Page 16: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

16

Experiments 2/6

Testing

BIKE !

Page 17: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

17

Experiments 3/6

Testing

BIKE ! BIKE !

Page 18: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

18

Experiments 4/6

Testing

NO BIKE ! NO BIKE !

Page 19: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

19

Experiments 5/6

Testing

NO BIKE !

Page 20: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

20

Experiments 6/6

Facts:

Page 21: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

21

Discussion / Outlook

• Further Experimental Evaluation

• Multiclass Categorisation

• Combination with other Types of Regions

Page 22: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

22

Conclusion

• Generic object recognition

• A new Framework

• A new Learning Model

• Good Results

Page 23: A. Opelt, M. Fussenegger, A. Pinz, P. Auer

Andreas Opelt (Graz University of Technology and University of Leoben)

23

Thank you !

Generic object recognition; not an easy task!

Thanks to the Lava Project and the FWF Project – FSP Cognitive Vision