evaluation of color descriptors for object and scene

15
Evaluation of Color Descriptors for Object and Scene Recognition Authors: Koen E. A. van de Sande, Theo Gevers, and Cees G. M. Snoek @ University of Amsterdam Presenter: Shao-Chuan Wang

Upload: shao-chuan-wang

Post on 24-May-2015

2.450 views

Category:

Technology


0 download

TRANSCRIPT

Page 1: Evaluation Of Color Descriptors For Object And Scene

Evaluation of Color Descriptors for Object and Scene Recognition

Authors: Koen E. A. van de Sande, Theo Gevers, and Cees G. M. Snoek @

University of AmsterdamPresenter: Shao-Chuan Wang

Page 2: Evaluation Of Color Descriptors For Object And Scene

Evaluation of Color Descriptors for Object and Scene Recognition

• Focus: Color features/descriptors on obj. and scene recognition

• Summary: The invariance of photometric transform aces and its effect on discriminative power.

• Conclusion: The usefulness of invariance is category-specific!

Page 3: Evaluation Of Color Descriptors For Object And Scene

Photometric transforms (1/2)

• Light intensity scale invariant

• Light intensity shift invariant

B

G

R

a

a

a

00

00

00

3

2

1

o

o

o

B

G

R

Page 4: Evaluation Of Color Descriptors For Object And Scene

Photometric transforms (1/2)

• Light intensity scale and shift invariant

• Light color change

• Light color change and shift

3

2

1

00

00

00

o

o

o

B

G

R

a

a

a

B

G

R

c

b

a

00

00

00

3

2

1

00

00

00

o

o

o

B

G

R

c

b

a

Page 5: Evaluation Of Color Descriptors For Object And Scene

Color Descriptors (1/1)

• Histograms – RGB, Hue, Saturation, rgHistogram, …

• Color Moments: contain spatial info.

• Color SIFT: combined color and SIFT– HSV-SIFT, Hue-SIFT, …

dxdyyxIyxIyxIyxM cB

bG

aR

qpabcpq )],([)],([)],([

Page 6: Evaluation Of Color Descriptors For Object And Scene

Color Histograms (1/2)

• RGB-histogram• Hue-histogram– H and S are scale-invariant and

shift-invariant w.r.t light intensity

• rg-histogram– Scale-invariant– Not shift-invariant– Not that b is redudant

BGRB

BGRG

BGRR

b

g

r

Image from wikipedia

Page 7: Evaluation Of Color Descriptors For Object And Scene

Color Histograms (2/2)

• Transformed color– Scale and shift-invariant w.r.t

light intensity.

• Opponent color histogram– O1,O2 shift invariant– O3: intensity, no invariant

B

B

G

G

R

R

B

G

R

B

G

R

3

622

3

2

1

BGR

BGR

GR

O

O

O

Page 8: Evaluation Of Color Descriptors For Object And Scene

Color SIFT Descriptors (1/3)

• HSV-SIFT– SIFT over HSV channels– Hue is unstable in gray axis

• Hue-SIFT (Van de Weijer 2006)

– Used Hue histogram weighing by its saturation

– Concatenate Hue histogram with SIFT– Only SIFT is invariant; Hue histogram is not!

(partial invariance)

Hue Instability

Page 9: Evaluation Of Color Descriptors For Object And Scene

Color SIFT Descriptors (2/3)

• OpponentSIFT– SIFT over all channels in the opponent color space.– Shift-invariant to light intensity.

• W-SIFT– Eliminate O1 and O2’s intensity information– Scale-invariant to light intensity

• rg-SIFT– SIFT over r,g spaces– Scale and shift invariant, but not invariant to light color

changes/shifts

32

31

2

1

OO

OO

O

O

Page 10: Evaluation Of Color Descriptors For Object And Scene

Color SIFT Descriptors (3/3)

• Transformed color SIFT– SIFT over normalized transformed channels.– Scale- and shift-invariant to light color changes

and shift.

Page 11: Evaluation Of Color Descriptors For Object And Scene

Experiments

• Implementation:– Scale-invariants points are detected by Harris-

Laplace point detectors– Color descriptors are computed over the area

around the points; all regions are proportionally re-sampled to a uniform 60x60 patch.

– Cluster descriptors with k = 40 (images) k = 4000 (video)

– SVM classifier with EMD/chi-square kernel

Page 12: Evaluation Of Color Descriptors For Object And Scene

Benchmark (1/3)

• Image: PASCAL VOC 2007, over 20 object categories

Page 13: Evaluation Of Color Descriptors For Object And Scene

Benchmark (2/3)

• Most objs were categorized better under scale- and shift- invariant to light intensity

• Some, such as car and dining table, do not benefit from such invariance.

Page 14: Evaluation Of Color Descriptors For Object And Scene

Benchmark (3/3)

• Video: Mediamill Challenge, 39 object and scene categories

Page 15: Evaluation Of Color Descriptors For Object And Scene

Evaluation of Color Descriptors for Object and Scene Recognition

• Conclusion:– W-SIFT and rgSIFT, in general, outperform other

color descriptors.– Light intensity info. Is important for some

categories– Usefulness of invariance is category-specific.