SIFT The SIFT (Scale Invariant Feature Transform) Detector and Descriptor developed by David Lowe University of British Columbia Initial paper IC

The SIFT (Scale Invariant Feature Transform) Detector and Descriptor 

developed by David Lowe

University of British Columbia

Initial paper ICCV 1999

Newer journal paper IJCV 2004

 
 

2  

Review: Matt Brown��s Canonical Frames

 
 

3  

Multi-Scale Oriented Patches 

• Extract oriented patches at multiple scales

 

[ Brown, Szeliski, Winder CVPR 2005 ]

 
 

4  

Application: Image Stitching 

[ Microsoft Digital Image Pro version 10 ]

Ideas from Matt��s Multi-Scale Oriented Patches  

• 1. Detect an interesting patch with an interest operator. Patches are translation invariant.
• 2. Determine its dominant orientation.
• 3. Rotate the patch so that the dominant orientation points upward. This makes the patches rotation invariant.
• 4. Do this at multiple scales, converting them all to one scale through sampling.
• 5. Convert to illumination ��invariant�� form

 
 
 

5

 
 

6  

Implementation Concern: 
How do you rotate a patch? 

• Start with an ��empty�� patch whose dominant direction is ��up��.
• For each pixel in your patch, compute the position in the detected image patch. It will be in floating point and will fall between the image pixels.
• Interpolate the values of the 4 closest pixels in the image, to get a value for the pixel in your patch.

 
 

7  

Rotating a Patch 

empty canonical patch 

patch detected in the image 

x�� = x cos�� – y sin��

y�� = x sin�� + y cos�� 

T 

T 

counterclockwise rotation 

(x,y) 

(x��,y��)

 
 

8  

Using Bilinear Interpolation 

• Use all 4 adjacent samples

 

x 

y 

I₀₀ 

I₁₀ 

I₀₁ 

I₁₁

 
 

9  

SIFT: Motivation 

• The Harris operator is not invariant to scale and  correlation is not invariant to rotation^1.

• For better image matching, Lowe��s goal was to develop an interest operator that is invariant to scale and rotation. 
  

• Also, Lowe aimed to create a descriptor that was robust to the variations corresponding to typical viewing conditions. The descriptor is the most-used part of SIFT. 
  

 
 
 

¹But Schmid and Mohr developed a rotation invariant descriptor for it in 1997.

 
 

10  

Idea of SIFT  

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

 

SIFT Features

 
 

11  

Claimed Advantages of SIFT   

• Locality: features are local, so robust to occlusion and clutter (no prior segmentation)
• Distinctiveness: individual features can be matched to a large database of objects
• Quantity: many features can be generated for even small objects
• Efficiency: close to real-time performance
• Extensibility: can easily be extended to wide range of differing feature types, with each adding robustness

 
 

12  

Overall Procedure at a High Level 

1. Scale-space extrema detection

1. Keypoint localization 
    
   

1. Orientation assignment 
    
   

1. Keypoint description 
    
   

 

Search over multiple scales and image locations. 

Fit a model to detrmine location and scale.

Select keypoints based on a measure of stability. 

Compute best orientation(s) for each keypoint region. 

Use local image gradients at selected scale and rotation

to describe each keypoint region.

 
 

13  

1. Scale-space extrema detection 

• Goal: Identify locations and scales that can be repeatably assigned under different views of the same scene or object.
• Method: search for stable features across multiple scales using a continuous function of scale.
• Prior work has shown that under a variety of assumptions, the best function is a Gaussian function.
• The scale space of an image is a function L(x,y,) that is produced from the convolution of a Gaussian kernel (at different scales) with the input image.

 
 

14  

Aside: Image Pyramids 

Bottom level is the original image. 

2^nd level is derived from the

original image according to

some function 

3^rd level is derived from the

2^nd level according to the same

funtion 

And so on.

 
 

15  

Aside: Mean Pyramid 

Bottom level is the original image. 

At 2^nd level, each pixel is the mean

of 4 pixels in the original image. 

At 3^rd level, each pixel is the mean

of 4 pixels in the 2^nd level. 

And so on. 

mean

 
 

16  

Aside: Gaussian Pyramid 
At each level, image is smoothed and reduced in size. 

Bottom level is the original image. 

At 2^nd level, each pixel is the result

of applying a Gaussian mask to

the first level and then subsampling

to reduce the size. 

And so on. 

Apply Gaussian filter

 
 

17  

Example: Subsampling with Gaussian pre-filtering 

G 1/4 

G 1/8 

Gaussian 1/2 

 

 
 

18  

Lowe��s Scale-space Interest Points 

• Laplacian of Gaussian kernel
• Scale normalised (x by scale²)
• Proposed by Lindeberg
• Scale-space detection
• Find local maxima across scale/space
• A good ��blob�� detector

 
 
 

[ T. Lindeberg IJCV 1998 ]

 
 

19  

Lowe��s Scale-space Interest Points: 
Difference of Gaussians 

• Gaussian is an ad hoc solution of heat diffusion equation

• Hence 
   
  

• k is not necessarily very small in practice 
   
  

 
 

20  

Lowe��s Pyramid Scheme 

• Scale space is separated into octaves:
• Octave 1 uses scale 
• Octave 2 uses scale 2
• etc.

• In each octave, the initial image is repeatedly convolved 
  

  with Gaussians to produce a set of  scale space images. 

• Adjacent Gaussians are subtracted to produce the DOG

• After each octave, the Gaussian image is down-sampled 
  

   by a factor of 2 to produce an image ¼ the size to start

   the next level.

 
 

21  

Lowe��s Pyramid Scheme 

s+2 filters

_s+1=2^(s+1)/s₀ 

.

.

_i=2^i/s₀

.

.

₂=2^2/s₀

₁=2^1/s₀

₀ 

s+3

images

including

original 

s+2

differ-

ence

images 

The parameter s determines the number of images per octave.

 
 

22  

Key point localization 

• Detect maxima and minima of difference-of-Gaussian in scale space

• Each point is compared to its 8 neighbors in the current image and 9 neighbors each in the scales above and below 
  

 

For each max or min found,

output is the location and

the scale. 

s+2 difference images.

top and bottom ignored.

s planes searched.

 
 

23  

Scale-space extrema detection: experimental results over 32 images that were synthetically transformed and noise added. 

• Sampling in scale for efficiency
• How many scales should be used per octave? S=?
• More scales evaluated, more keypoints found
• S < 3, stable keypoints increased too
• S > 3, stable keypoints decreased
• S = 3, maximum stable keypoints found

 

% detected 

           % correctly matched 

average no. detected 

average no. matched 

Stability 

Expense

 
 

24  

Keypoint localization  

• Once a keypoint candidate is found, perform a detailed fit to nearby data to determine
• location, scale, and ratio of principal curvatures
• In initial work keypoints were found at location and scale of a central sample point.
• In newer work, they fit a 3D quadratic function to improve interpolation accuracy.
• The Hessian matrix was used to eliminate edge responses.

 

 
 

25  

Eliminating the Edge Response 

• Reject flats:
•                     < 0.03
• Reject edges:

• r < 10 
   
   
   
   
   
  
• What does this look like?

 

Let  be the eigenvalue with

larger magnitude and  the smaller. 

Let r = /.

So  = r 

(r+1)²/r  is at a

min when the

2 eigenvalues

are equal.

 
 

26  

3. Orientation assignment 

• Create histogram of local gradient directions at selected scale
• Assign canonical orientation at peak of smoothed histogram
• Each key specifies stable 2D coordinates (x, y, scale,orientation)

 
 

If 2 major orientations, use both.

 
 

27  

Keypoint localization with orientation 

832 

729 

536 

233x189 

initial keypoints 

keypoints after

gradient threshold 

keypoints after

ratio threshold

 
 

28  

4. Keypoint Descriptors 

• At this point, each keypoint has
• location
• scale
• orientation
• Next is to compute a descriptor for the local image region about each keypoint that is
• highly distinctive
• invariant as possible to variations such as changes in viewpoint and illumination

 
 

29  

Normalization 

• Rotate the window to standard orientation

• Scale the window size based on the scale at which the point was found. 
  

 
 

30  

Lowe��s Keypoint Descriptor 
(shown with 2 X 2 descriptors over 8 X 8) 

In experiments, 4x4 arrays of 8 bin histogram is used,

a total of 128 features for one keypoint

 
 

31  

Lowe��s Keypoint Descriptor 

• use the normalized region about the keypoint
• compute gradient magnitude and orientation at each point in the region
• weight them by a Gaussian window overlaid on the circle
• create an orientation histogram over the 4 X 4 subregions of the window
• 4 X 4 descriptors over 16 X 16 sample array were used in practice. 4 X 4 times 8 directions gives a vector of 128 values.

 

...

 
 

32  

Using SIFT for Matching ��Objects��

 
 

33

 
 

34  

Uses for SIFT  

• Feature points are used also for:
• Image alignment (homography, fundamental matrix)
• 3D reconstruction (e.g. Photo Tourism)
• Motion tracking
• Object recognition
• Indexing and database retrieval
• Robot navigation
• �� many others

 

[ Photo Tourism: Snavely et al. SIGGRAPH 2006 ]