Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen-...

34
Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society Conference on Computer Vision and Pattern Recognition(2005) vol. 1 ,pp.886 - 893

Transcript of Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen-...

Page 1: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Histograms of Oriented Gradients for Human Detection(HOG)

Presenter :JIA-HONG,DONG

Advisor : Yen- Ting, Chen

1

Dalal, N.; Triggs, B., IEEE Computer Society Conference on Computer Vision and Pattern Recognition(2005) vol. 1 ,pp.886 - 893  

Page 2: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Outline1. Introduction2. Methodology3. Results4. Discussion 5. Conclusion

2

Page 3: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Introduction Detecting humans in images is a

challenging task Variable appearance Wide range of poses

A robust feature set Discriminate cleanly

Cluttered backgrounds Different illumination

3

Page 4: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Introduction Edge orientation histograms

Scale-invariant feature transform (SIFT) Shape context

4

Shape contextSIFT

Page 5: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Introduction Using linear SVM as a baseline classifier Using detection error tradeoff (DET) Data Sets

MIT pedestrian set INRIA pedestrian set

5

Page 6: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Methodology Data Sets

MIT pedestrian database 509 training images 200 test images

INRIA 1805 64X128 images

6

Page 7: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Methodology

7

1 2 3 4 5

6 7 8 9 10

Page 8: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Methodology Training examples 12180+ examples

2478 Positive 1218 Negative

8

Page 9: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Methodology Detection error tradeoff

X-axes False Positives Per Window tested(by 5% at 10-4) FPPW=

Y-axes Miss rate=

Log-log scale

9

Page 10: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Methodology

10

Page 11: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Gamma/Color Normalization Inputting pixel representations

Grayscale RGB color spaces LAB color spaces

Power law (Gamma equalization)

11

Page 12: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

LAB Color Spaces

12

 Xn, Yn and Zn are the CIE XYZ tristimulus values of the reference white point

Page 13: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Power Law (Gamma equalization) Tradition

13

IGray(i, j) is the gray-level imageIEq (i, j) is the image which performed equalization IMax and IMin are the maximum and minimum of the pixel values of IGray(i, j)

Page 14: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Power Law (Gamma equalization)

14

i is the i-th gray levelL is the low-boundR is the actual equalization rangeGE (i) is the result of the i-th gray level obtained from gamma equalization

Page 15: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Power Law (Gamma equalization)

15

Page 16: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Gradient Computation Masks test(for each color channel)

Gaussian (σ=0~3) 1-D point derivatives[-1,0,1] Cubic-corrected[1,-8,0,8,-1] 3X3 Sobel mask

2X2 diagonal ones

16

Page 17: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Gradient Computation

17

‘c-cor’ is the 1D cubic-correctedpoint derivative

Page 18: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Spatial / Orientation Binning Orientation bins are evenly spaced

0 °~180 ° 0 °~360 °

18

Page 19: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Spatial / Orientation Binning

19

Page 20: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Normalization and Descriptor Blocks

20

Page 21: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Normalization and Descriptor Blocks

21

Page 22: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Normalization and Descriptor Blocks

22

Block Normalization schemes (limiting the maximum values of

v to 0.2) and renormalizing

Centre-surround normalization Window norm(using Gaussian σ=1)

v is the unnormalized descriptor vector is a small constant

Page 23: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Normalization and Descriptor Blocks

23

Page 24: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Normalization and Descriptor Blocks

24

Illumination and foreground-background contrast overlap

Page 25: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Normalization and Descriptor Blocks

25

Page 26: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Detector Window and Context

26

Page 27: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Classifier Using linear SVM(Support vector machine) Increasing performance

Using a Gaussian kernel Higher run time

27

Page 28: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Classifier

28

Using a Gaussian kernel SVM,

Page 29: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Results

29

Page 30: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Results

30

Page 31: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Results

31

The performance of selected detectors on (left) MIT and (right) INRIA data sets.

Page 32: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Discussion HOG outperform wavelet & shape context Traditional centre-surround style schemes

are not the best choice Similar to SIFT descriptors

32

Page 33: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Conclusion Scale gradients Orientation binning Relatively coarse spatial binning High-quality local contrast normalization in

overlapping descriptor blocks

33

Page 34: Histograms of Oriented Gradients for Human Detection(HOG) Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 Dalal, N.; Triggs, B., IEEE Computer Society.

Thank you for your attention

34