Stanford CS223B Computer Vision, Winter 2005 Lecture 4 Advanced Features

1
Stanford CS223B Computer Vision, Winter
2005Lecture 4 Advanced Features

• Sebastian Thrun, Stanford
• Rick Szeliski, Microsoft
• Hendrik Dahlkamp, Stanford
• with slides by D Lowe, M. Polleyfeys

2
Todays Goals

• Harris Corner Detector
• Hough Transform
• Templates, Image Pyramid, Transforms
• SIFT Features

3
Finding Corners
Edge detectors perform poorly at corners. Corners
provide repeatable points for matching, so are
worth detecting.

• Idea
• Exactly at a corner, gradient is ill defined.
• However, in the region around a corner, gradient
  has two or more different values.

4
The Harris corner detector
Form the second-moment matrix
Gradient with respect to x, times gradient with
respect to y
Sum over a small region around the hypothetical
corner
Matrix is symmetric
Slide credit David Jacobs
5
Simple Case
First, consider case where
This means dominant gradient directions align
with x or y axis If either ? is close to 0, then
this is not a corner, so look for locations where
both are large.
Slide credit David Jacobs
6
General Case
It can be shown that since C is symmetric
So every case is like a rotated version of the
one on last slide.
Slide credit David Jacobs
7
So, To Detect Corners

• Filter image with Gaussian to reduce noise
• Compute magnitude of the x and y gradients at
  each pixel
• Construct C in a window around each pixel (Harris
  uses a Gaussian window just blur)
• Solve for product of ls (determinant of C)
• If ls are both big (product reaches local maximum
  and is above threshold), we have a corner (Harris
  also checks that ratio of ls is not too high)

8
Gradient Orientation
9
Corner Detection
Corners are detected where the product of the
ellipse axis lengths reaches a local maximum.
10
Harris Corners

• Originally developed as features for motion
  tracking
• Greatly reduces amount of computation compared to
  tracking every pixel
• Translation and rotation invariant (but not scale
  invariant)

11
Harris Corner in Matlab

• Harris Corner detector - by Kashif Shahzad
• sigma2 thresh0.1 sze11 disp0
• Derivative masks
• dy -1 0 1 -1 0 1 -1 0 1
• dx dy' dx is the transpose matrix of dy
• Ix and Iy are the horizontal and vertical edges
  of image
• Ix conv2(bw, dx, 'same')
• Iy conv2(bw, dy, 'same')
• Calculating the gradient of the image Ix and Iy
• g fspecial('gaussian',max(1,fix(6sigma)),
  sigma)
• Ix2 conv2(Ix.2, g, 'same') Smoothed squared
  image derivatives
• Iy2 conv2(Iy.2, g, 'same')
• Ixy conv2(Ix.Iy, g, 'same')
• My preferred measure according to research
  paper
• cornerness (Ix2.Iy2 - Ixy.2)./(Ix2 Iy2
  eps)

12
Todays Goals

• Harris Corner Detector
• Hough Transform
• Templates, Image Pyramid, Transforms
• SIFT Features

13
Features?
Local versus global
14
Vanishing Points
15
Vanishing Points
A. Canaletto 1740, Arrival of the French
Ambassador in Venice
16
From Edges to Lines
17
Hough Transform
18
Hough Transform Quantization
m
Detecting Lines by finding maxima / clustering in
parameter space
19
Hough Transform Results
Hough Transform
Image
Edge detection
20
Todays Goals

• Harris Corner Detector
• Hough Transform
• Templates, Image Pyramid, Transforms
• SIFT Features

21
Problem Features for Recognition
in here
Want to find
22
Convolution with Templates

• read image
• im imread('bridge.jpg')
• bw double(im(,,1)) ./ 256
• imshow(bw)
• apply FFT
• FFTim fft2(bw)
• bw2 real(ifft2(FFTim))
• imshow(bw2)
• define a kernel
• kernelzeros(size(bw))
• kernel(1, 1) 1
• kernel(1, 2) -1
• FFTkernel fft2(kernel)
• apply the kernel and check out the result
• FFTresult FFTim . FFTkernel
• result real(ifft2(FFTresult))

• select an image patch
• patch bw(221240,351370)
• imshow(patch)
• patch patch - (sum(sum(patch)) / size(patch,1)
  / size(patch, 2))
• kernelzeros(size(bw))
• kernel(1size(patch,1),1size(patch,2)) patch
• FFTkernel fft2(kernel)
• apply the kernel and check out the result
• FFTresult FFTim . FFTkernel
• result max(0, real(ifft2(FFTresult)))
• result result ./ max(max(result))
• result (result . 1 gt 0.5)
• imshow(result)
• alternative convolution
• imshow(conv2(bw, patch, 'same'))

23
Convolution with Templates

• Invariances
• Scaling
• Rotation
• Illumination
• Deformation
• Provides
• Good localization

24
Scale Invariance Image Pyramid
25
Aliasing
26
Aliasing Effects
27
Convolution Theorem
Fourier Transform of g
F is invertible
28
The Fourier Transform

• Represent function on a new basis
• Think of functions as vectors, with many
  components
• We now apply a linear transformation to transform
  the basis
• dot product with each basis element
• In the expression, u and v select the basis
  element, so a function of x and y becomes a
  function of u and v
• basis elements have the form

29

• Fourier basis element
• example, real part
• Fu,v(x,y)
• Fu,v(x,y)const. for (uxvy)const.
• Vector (u,v)
• Magnitude gives frequency
• Direction gives orientation.

30
Here u and v are larger than in the previous
slide.
31
And larger still...
32
Phase and Magnitude

• Fourier transform of a real function is complex
• difficult to plot, visualize
• instead, we can think of the phase and magnitude
  of the transform
• Phase is the phase of the complex transform
• Magnitude is the magnitude of the complex
  transform
• Curious fact
• all natural images have about the same magnitude
  transform
• hence, phase seems to matter, but magnitude
  largely doesnt
• Demonstration
• Take two pictures, swap the phase transforms,
  compute the inverse - what does the result look
  like?

33
(No Transcript)
34
This is the magnitude transform of the cheetah
picture
35
This is the phase transform of the cheetah picture
36
(No Transcript)
37
This is the magnitude transform of the zebra
picture
38
This is the phase transform of the zebra picture
39
Reconstruction with zebra phase, cheetah magnitude
40
Reconstruction with cheetah phase, zebra magnitude
41
Convolution with Templates

• Invariances
• Scaling
• Rotation
• Illumination
• Deformation
• Provides
• Good localization

42
Todays Goals

• Harris Corner Detector
• Hough Transform
• Templates, Image Pyramid, Transforms
• SIFT Features

43
Lets Return to this Problem
in here
Want to find
44
SIFT

• Invariances
• Scaling
• Rotation
• Illumination
• Deformation
• Provides
• Good localization

45
SIFT Reference

• Distinctive image features from scale-invariant
  keypoints. David G. Lowe, International Journal
  of Computer Vision, 60, 2 (2004), pp. 91-110.
• SIFT Scale Invariant Feature Transform

46
Invariant Local Features

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

SIFT Features
47
Advantages of invariant local features

• 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

48
SIFT On-A-Slide

• Enforce invariance to scale Compute Gaussian
  difference max, for may different scales
  non-maximum suppression, find local maxima
  keypoint candidates
• Localizable corner For each maximum fit
  quadratic function. Compute center with sub-pixel
  accuracy by setting first derivative to zero.
• Eliminate edges Compute ratio of eigenvalues,
  drop keypoints for which this ratio is larger
  than a threshold.
• Enforce invariance to orientation Compute
  orientation, to achieve scale invariance, by
  finding the strongest second derivative direction
  in the smoothed image (possibly multiple
  orientations). Rotate patch so that orientation
  points up.
• Compute feature signature Compute a "gradient
  histogram" of the local image region in a 4x4
  pixel region. Do this for 4x4 regions of that
  size. Orient so that largest gradient points up
  (possibly multiple solutions). Result feature
  vector with 128 values (15 fields, 8 gradients).
• Enforce invariance to illumination change and
  camera saturation Normalize to unit length to
  increase invariance to illumination. Then
  threshold all gradients, to become invariant to
  camera saturation.

49
SIFT On-A-Slide

• Enforce invariance to scale Compute Gaussian
  difference max, for may different scales
  non-maximum suppression, find local maxima
  keypoint candidates
• Localizable corner For each maximum fit
  quadratic function. Compute center with sub-pixel
  accuracy by setting first derivative to zero.
• Eliminate edges Compute ratio of eigenvalues,
  drop keypoints for which this ratio is larger
  than a threshold.
• Enforce invariance to orientation Compute
  orientation, to achieve scale invariance, by
  finding the strongest second derivative direction
  in the smoothed image (possibly multiple
  orientations). Rotate patch so that orientation
  points up.
• Compute feature signature Compute a "gradient
  histogram" of the local image region in a 4x4
  pixel region. Do this for 4x4 regions of that
  size. Orient so that largest gradient points up
  (possibly multiple solutions). Result feature
  vector with 128 values (15 fields, 8 gradients).
• Enforce invariance to illumination change and
  camera saturation Normalize to unit length to
  increase invariance to illumination. Then
  threshold all gradients, to become invariant to
  camera saturation.

50
Find Invariant Corners

• Enforce invariance to scale Compute Gaussian
  difference max, for may different scales
  non-maximum suppression, find local maxima
  keypoint candidates

51
Finding Keypoints (Corners)

• Idea Find Corners, but scale invariance
• Approach
• Run linear filter (diff of Gaussians)
• At different resolutions of image pyramid

52
Difference of Gaussians
Minus
Equals
53
Difference of Gaussians

• surf(fspecial('gaussian',40,4))
• surf(fspecial('gaussian',40,8))
• surf(fspecial('gaussian',40,8) -
  fspecial('gaussian',40,4))

54
Find Corners with DiffOfGauss

• im imread('bridge.jpg')
• bw double(im(,,1)) / 256
• for i 1 10
• gaussD fspecial('gaussian',40,2i) -
  fspecial('gaussian',40,i)
• mesh(gaussD) drawnow
• res abs(conv2(bw, gaussD, 'same'))
• res res / max(max(res))
• imshow(res) title('\bf i ' num2str(i))
  drawnow
• end

55
Build Scale-Space Pyramid

• All scales must be examined to identify
  scale-invariant features
• An efficient function is to compute the
  Difference of Gaussian (DOG) pyramid (Burt
  Adelson, 1983)

56
Key point localization

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

57
Example of keypoint detection
(a) 233x189 image (b) 832 DOG extrema
58
SIFT On-A-Slide

• Enforce invariance to scale Compute Gaussian
  difference max, for may different scales
  non-maximum suppression, find local maxima
  keypoint candidates
• Localizable corner For each maximum fit
  quadratic function. Compute center with sub-pixel
  accuracy by setting first derivative to zero.
• Eliminate edges Compute ratio of eigenvalues,
  drop keypoints for which this ratio is larger
  than a threshold.
• Enforce invariance to orientation Compute
  orientation, to achieve scale invariance, by
  finding the strongest second derivative direction
  in the smoothed image (possibly multiple
  orientations). Rotate patch so that orientation
  points up.
• Compute feature signature Compute a "gradient
  histogram" of the local image region in a 4x4
  pixel region. Do this for 4x4 regions of that
  size. Orient so that largest gradient points up
  (possibly multiple solutions). Result feature
  vector with 128 values (15 fields, 8 gradients).
• Enforce invariance to illumination change and
  camera saturation Normalize to unit length to
  increase invariance to illumination. Then
  threshold all gradients, to become invariant to
  camera saturation.

59
Example of keypoint detection
Threshold on value at DOG peak and on ratio of
principle curvatures (Harris approach)

• (c) 729 left after peak value threshold
• (d) 536 left after testing ratio of principle
  curvatures

60
SIFT On-A-Slide

• Enforce invariance to scale Compute Gaussian
  difference max, for may different scales
  non-maximum suppression, find local maxima
  keypoint candidates
• Localizable corner For each maximum fit
  quadratic function. Compute center with sub-pixel
  accuracy by setting first derivative to zero.
• Eliminate edges Compute ratio of eigenvalues,
  drop keypoints for which this ratio is larger
  than a threshold.
• Enforce invariance to orientation Compute
  orientation, to achieve scale invariance, by
  finding the strongest second derivative direction
  in the smoothed image (possibly multiple
  orientations). Rotate patch so that orientation
  points up.
• Compute feature signature Compute a "gradient
  histogram" of the local image region in a 4x4
  pixel region. Do this for 4x4 regions of that
  size. Orient so that largest gradient points up
  (possibly multiple solutions). Result feature
  vector with 128 values (15 fields, 8 gradients).
• Enforce invariance to illumination change and
  camera saturation Normalize to unit length to
  increase invariance to illumination. Then
  threshold all gradients, to become invariant to
  camera saturation.

61
Select canonical 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)

62
SIFT On-A-Slide

• Enforce invariance to scale Compute Gaussian
  difference max, for may different scales
  non-maximum suppression, find local maxima
  keypoint candidates
• Localizable corner For each maximum fit
  quadratic function. Compute center with sub-pixel
  accuracy by setting first derivative to zero.
• Eliminate edges Compute ratio of eigenvalues,
  drop keypoints for which this ratio is larger
  than a threshold.
• Enforce invariance to orientation Compute
  orientation, to achieve scale invariance, by
  finding the strongest second derivative direction
  in the smoothed image (possibly multiple
  orientations). Rotate patch so that orientation
  points up.
• Compute feature signature Compute a "gradient
  histogram" of the local image region in a 4x4
  pixel region. Do this for 4x4 regions of that
  size. Orient so that largest gradient points up
  (possibly multiple solutions). Result feature
  vector with 128 values (15 fields, 8 gradients).
• Enforce invariance to illumination change and
  camera saturation Normalize to unit length to
  increase invariance to illumination. Then
  threshold all gradients, to become invariant to
  camera saturation.

63
SIFT vector formation

• 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

64
Nearest-neighbor matching to feature database

• Hypotheses are generated by approximate nearest
  neighbor matching of each feature to vectors in
  the database
• SIFT use best-bin-first (Beis Lowe, 97)
  modification to k-d tree algorithm
• Use heap data structure to identify bins in order
  by their distance from query point
• Result Can give speedup by factor of 1000 while
  finding nearest neighbor (of interest) 95 of the
  time

65
3D Object Recognition

• Extract outlines with background subtraction

66
3D Object Recognition

• Only 3 keys are needed for recognition, so extra
  keys provide robustness
• Affine model is no longer as accurate

67
Recognition under occlusion
68
Test of illumination invariance

• Same image under differing illumination

273 keys verified in final match
69
Examples of view interpolation
70
Location recognition
71

• Sony Aibo
• (Evolution Robotics)
• SIFT usage
• Recognize
• charging
• station
• Communicate
• with visual
• cards

72
Todays Goals

• Harris Corner Detector
• Hough Transform
• Templates, Image Pyramid, Transforms
• SIFT Features

This is the end of features....