Distinctive Image Features from Scale-Invariant Keypoints Presenter :JIA-HONG,DONG Advisor : Yen- Ting, Chen 1 David G. Lowe International Journal of Computer Vision,Volume 60 Issue 2, Pages 91 – 110, 2004.

Outline  Introduction  Methodology  Recognition examples  Simulation and testing  Conclusion 2

Introduction  Scale Invariant Feature Transform(SIFT) Object and scene recognition Video Tracking Robotic mapping and navigation Image stitching 3D modeling Gesture recognition Match moving 3

Introduction  Advantage The features are invariant:  Image scaling  Image rotation  Illumination  Noise  Camera viewpoint High performance:  High accuracy  Near real-time 4

Introduction  Major stages of computation: Scale-space extrema detection Keypoint localization Orientation assignment Keypoint descriptor Matching features(nearest neighbor) 5

Methodology  Detection of scale-space extrema Convolution image subtraction (1).....(2).....(3)

Methodology 7

(К-1) is a constant factor σ 2 is scale invariant as studied by Lindeberg (1994) σ 2 Δ 2 G is maxima and minima image features as studied by Mikolajczyk(2002) (4).....(5).....(6)

Methodology 9

 Database 32 real images:  Outdoor scenes  Human faces  Aerial photographs  Industrial images 10

Methodology  Database Transformations:  Rotation  Scaling  Brightness  Contrast  Noise 11

Methodology  Local extrema detection 12

Methodology  Local extrema detection 13

Methodology  Frequency of sampling in scale 14

Methodology  Accurate keypoint localization Using Taylor expansion x=(x, y, σ ) T is the offset from this point (3).....(8).....(9)

Methodology (10)

Methodology 17

Methodology 18

Methodology  Orientation assignment L(x,y) is the sample image. Θ(x,y) is orientation m(x,y) is the gradient magnitude 19...(11).....(12)

Methodology  Orientation histogram A region around the keypoint 36 bins covering the 360 degree range of orientations  Added weighted A Gaussian window σ=1.5 20

Methodology  The local image descriptor 21

Methodology  Descriptor testing 22

Methodology  Descriptor representation 4x4 array of histograms 8 orientation bins 4x4x8 = 128 element feature vector 23

Methodology  Keypoint matching Minimum Euclidean distance 24

Methodology  Keypoint matching 25

Methodology  Application to object recognition Need 3 features at least  Higher probability 26

Recognition examples  Descriptor testing 27

Recognition examples  Sensitivity to affine change 28

Recognition examples  Matching to large databases 29

Recognition examples 30

Recognition examples 31

Simulation and testing 32

Simulation and testing 33

Conclusion  SIFT keypoints described are particularly useful.  A high-dimensional vector represents the image gradients within a local region of the image.  Near real-time performance on standard PC hardware. 34

Conclusion  Systematic testing is needed on data sets with full 3D viewpoint.  Feature sets are likely to contain both prior and learned features. 35

36 Thank you for your attention