Scribd
Upload
109 views24 pages

SIFT: Scale-Invariant Feature Detection

SIFT extracts keypoints at multiple scales and orientations that are invariant to translation, rotation, and scale changes. Keypoints are located as local maxima/minima of the Difference of Gaussian function. Each keypoint is assigned a histogram of gradient orientations as its signature. Keypoints can then be matched between images based on nearest neighbor distances between signatures. SIFT has proven effective for correspondence matching and has inspired other local feature detectors and descriptors.

Uploaded by

Jessy Le
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
109 views24 pages

SIFT: Scale-Invariant Feature Detection

SIFT extracts keypoints at multiple scales and orientations that are invariant to translation, rotation, and scale changes. Keypoints are located as local maxima/minima of the Difference of Gaussian function. Each keypoint is assigned a histogram of gradient orientations as its signature. Keypoints can then be matched between images based on nearest neighbor distances between signatures. SIFT has proven effective for correspondence matching and has inspired other local feature detectors and descriptors.

Uploaded by

Jessy Le
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd

Object Recognition from

Local Scale-Invariant
Features (SIFT)
David G. Lowe

Presented by David Lee


3/20/2006
Introduction

 Well engineered local descriptor


Introduction
 Image content is transformed into local feature
coordinates that are invariant to translation,
rotation, scale, and other imaging parameters

SIFT Features
Introduction
 Initially proposed for correspondence
matching
 Proven to be the most effective in such cases according to a recent
performance study by Mikolajczyk & Schmid (ICCV ’03)
Introduction
 Automatic Mosaicing

 [Link]
Introduction

 Now being used for general object class


recognition (e.g. 2005 Pascal challenge)

 Histogram of gradients
 Human detection, Dalal & Triggs CVPR ’05
Introduction

 SIFT in one sentence


 Histogram of gradients @ Harris-corner-like
 Extract features
 Find keypoints
 Scale, Location
 Orientation
 Create signature

 Match features
Finding Keypoints – Scale,
Location
 How do we choose scale?
Finding Keypoints – Scale,
Location
 Scale selection principle (T. Lindeberg ’94)
 In the absence of other evidence, assume that a scale level, at
which (possibly non-linear) combination of normalized derivatives
assumes a local maximum over scales, can be treated as
reflecting a characteristic length of a corresponding structure in
the data.

 Maxima/minima of Difference of Gaussian


Finding Keypoints – Scale,
Location # of scales/octave
=> empirically

Downsample

Find extrema
in 3D DoG space

Convolve with
Gaussian
Finding Keypoints – Scale,
Location
 Sub-pixel Localization
 Fit Trivariate quadratic to
find sub-pixel extrema

 Eliminating edges
 Similar to Harris corner detector
Finding Keypoints – Scale,
Location
 Key issue: Stability (Repeatability)

 Alternatives
 Multi-scale Harris corner detector
 Harris-Laplacian
 Kadir & Brady Saliency Detector
Recall Fei-fei’s pLSA paper
 …
 Uniform grid sampling
 Random sampling
** Important Note ** Their application was scene classification
NOT correspondence matching
Finding Keypoints – Scale,
Location

← Laplacian →
scale
 Harris-Laplacian1
Find local maximum of:
 Laplacian in scale
y
 Harris corner detector

in space (image ← Harris → x


coordinates)

• SIFT2 scale

← DoG →
Find local maximum of:
– Difference of y
Gaussians in space and
scale ← DoG → x
1
[Link], [Link]. “Indexing Based on Scale Invariant Interest Points”. ICCV 2001
2
[Link]. “Distinctive Image Features from Scale-Invariant Keypoints”. IJCV 2004
Finding Keypoints –
Orientation
 Create histogram of local
gradient directions
computed at selected scale
 Assign canonical
orientation at peak of
smoothed histogram
 Each key specifies stable
2D coordinates (x, y, scale,
orientation)

0 2π
Finding Keypoints –
Orientation
 Assign dominant
orientation as the
orientation of the
keypoint
Finding Keypoints

 So far, we found…
 where interesting things are happening
 and its orientation

 With the hope of


 Same keypoints being found, even under some
scale, rotation, illumination variation.
 Extract features
 Find keypoints
 Scale, Location
 Orientation
 Create signature

 Match features
Creating Signature
 Thresholded image gradients are sampled over
16x16 array of locations in scale space
 Create array of orientation histograms
 8 orientations x 4x4 histogram array = 128
dimensions
# dimension
=> empirically
Creating Signature

 What kind of information does this capture?


Comparison with HOG (Dalal
’05)
 Histogram of Oriented Gradients
 General object class recognition (Human)
 Engineered for a different goal

 Uniform sampling
 Larger cell (6-8 pixels)
 Fine orientation binning
 9 bins/180O vs. 8 bins/360O

 Both are well engineered


Comparison with MOPS
(Brown ’05)
 Multi-Image Matching using Multi-Scale
Orientated Patches (CVPR ’05)

 Simplified SIFT
 Multi-scale Harris corner
 No Histogram in orientation selection
 Smoothed image patch as descriptor

 Good performance for panorama stitching


 Extract features
 Find keypoints
 Scale, Location
 Orientation
 Create signature

 Match features
 Nearest neighbor, Hough voting, Least-square affine
parameter fit
Conclusion

 A novel method for detecting interest points

 Histogram of Oriented Gradients are


becoming more popular

 SIFT may not be optimal for general object


classification

You might also like