16
SIFT - Distinctive Image Features from Scale-Invariant Keypoints by David G. Lowe presented by David Störmer

SIFT - Distinctive Image Features from Scale-Invariant Keypoints

Embed Size (px)

DESCRIPTION

Videos:http://www.youtube.com/watch?v=NTGC6kqN244http://www.youtube.com/watch?v=GhMVf9HN3t0

Citation preview

Page 1: SIFT - Distinctive Image Features from Scale-Invariant Keypoints

SIFT - Distinctive Image Features from Scale-Invariant

Keypointsby David G. Lowe

presented by David Störmer

Page 2: SIFT - Distinctive Image Features from Scale-Invariant Keypoints

Table of contents

Feature Generation● detection of scale-space

extrema● local extrema detection● maxima/minima filtering

with thresholds● subpixel extrema

detection

Classification Generation● gradient

direction/magnitude diagramm on key-point

● Histogram of directions ● generate SIFT feature

vector

Page 3: SIFT - Distinctive Image Features from Scale-Invariant Keypoints

Detection of scale-space extrema

Page 4: SIFT - Distinctive Image Features from Scale-Invariant Keypoints

Local extrema detection

● search for maxima/minima within a 3x3x3 region

● performed on results of difference of Gaussian algorithm

Page 5: SIFT - Distinctive Image Features from Scale-Invariant Keypoints

Maxima/Minima filtering with thresholds

● in local area there a lots of values

● local extrema can have very small values

● double thresholding helps to remove "unimportant" points that were created from small difference in source image, such as noise

Page 6: SIFT - Distinctive Image Features from Scale-Invariant Keypoints

Subpixel extrema detection

● the given data does not represent a signal exactly, because it is sampled

● position of extrema can be extrapolated by using Tylor expression and setting it to zero

● the extrema can also get another value

● this subpixel extrema is done for direction x, y and can also be done for z (between DoG)

Page 7: SIFT - Distinctive Image Features from Scale-Invariant Keypoints

Gradient direction/magnitude

● Calculate gradient and magnitude of a 16x16 region around key-point

● create a kind of characteristic of keypoint in location

Page 8: SIFT - Distinctive Image Features from Scale-Invariant Keypoints

Histogram of directions

● directions are separated into 36 bins with each 10 degrees● magnitude of each direction is added on certain bin● each bar with a peak over 80% is converted into

a separate key-point with its own direction

Page 9: SIFT - Distinctive Image Features from Scale-Invariant Keypoints

Generate SIFT feature vector - 1

Aim of this step: generate 128 dimensional feature vector of 16x16 window around key-point

Page 10: SIFT - Distinctive Image Features from Scale-Invariant Keypoints

Generate SIFT feature vector - 2

● in 4x4 Window gradient an magnitude is calculated● foreach, the key-point orientation is subtracted● orientation were put in 8 bin histogramm

Page 11: SIFT - Distinctive Image Features from Scale-Invariant Keypoints

Generate SIFT feature vector - 2

● depending on their distance to key-point, a gaussian weighting function is applied

● the result: 8 bins with directions● this is done for alle 4x4 regions within 16x16 region = 128

directions

Page 12: SIFT - Distinctive Image Features from Scale-Invariant Keypoints

Results in my thesis

Page 13: SIFT - Distinctive Image Features from Scale-Invariant Keypoints

Results in my thesis

Page 14: SIFT - Distinctive Image Features from Scale-Invariant Keypoints

Results in my thesis

Page 15: SIFT - Distinctive Image Features from Scale-Invariant Keypoints

Thanks

Page 16: SIFT - Distinctive Image Features from Scale-Invariant Keypoints

Sources

● http://www.cs.ubc.ca/~lowe/keypoints/ ● http://en.wikipedia.org/wiki/Scale-

invariant_feature_transform ● http://www.aishack.in/2010/05/sift-scale-invariant-feature-

transform/