Visual Object Recognition Gutemberg Guerra

1
Visual Object RecognitionGutemberg Guerra

• Object Recognition from Local Scale-Invariant
  Features
• By David Lowe

2
Problem
3
Invariant but Distinctive

• Noise
• Illumination
• Affine
• Scale
• Rotation
• Translation
• Viewpoint
• 3D Projective

4
Scale Invariant Feature Transform (SIFT)

• Dense local features
• Local image descriptors
• Sampled at a large number of repeatable locations
• Transforms an image into a large collection of
  local feature vectors

5
Detection Stages

• Scale-space extrema detection
• Maxima or minima of a difference-of-Gaussian
  function
• Keypoint localization
• Identify 1,000 key locations in scale space
• Orientation assignment
• Generation of keypoint descriptors (feature
  vectors)
• Describe the local image region sample relative
  to its scale-space coordinate frame

6
Initialization

• Image smoothing
• Expand image by a factor of 2 using bilinear
  interpolation
• Results in 1,000 key points for a 512x512 pixel
  image

7
Gaussian function
8
Difference-of-Gaussian
9
Image Pyramid

• Resample image B using bilinear interpolation
  with pixel spacing of 1.5 in each direction

10
Gaussian Blurred Images
11
Difference of Gaussian Images
12
Local Extrema Detection

• Compare each pixel in the pyramid to its neighbors

13
Candidate Keypoint

• Position determined by Interpolation of nearby
  data
• Remove low contrast
• Eliminate responses along edges

14
Keypoint Removal
15
Final Keypoints
16
Orientation

• At each level of the pyramid
• Extract image gradients and orientation from the
  smoothed image A

17
Canonical Orientation

• Peak in a histogram of local image gradient
  orientations
• Histogram uses gradient magnitude and a Gaussian
  window to weight each neighboring pixel
  contribution
• Histogram uses 36 bins covering 360 degrees
• Separate keypoints for histogram maximum and
  directions with 80 of maximum
• Keypoint properties measured relative to
  orientation

18
Keys Detected

• Rotate 15 degrees
• Scale by a factor of 0.9
• Stretch by a factor of 1.1 in the horizontal
  direction
• Pixel intensities have 0.1 subtracted
• Contrast reduced by multiplication by 0.9
• Random pixel noise is added to give less than 5
  bits/pixel of signal
• 78 of keys match

19
Key Stability
20
Local Image Description

• Multiple images representing each of a number (8)
  of orientation planes sampled over a 4x4 grid of
  locations
• Pixels in a circle around the key location are
  inserted into the orientation planes
• Sample the image at a larger scale with a 2x2
  grid
• SIFT key vector 8x4x4 8x2x2 160

21
Feature Descriptor
22
Recognition Stages

• Compute SIFT features on the query image
• Match these features to the training database
  using a nearest-neighbor approach
• Use Hough transform to identify clusters of
  matches that vote for the same object pose
• Elements in the Hough accumulator that accumulate
  at least 3 votes are selected
• Least-squares fit to a final estimate of model
  parameters
• If 3 keys agree with low residual

23
Indexing and Feature Matching

• Find nearest neighbors in a database of SIFT
  features from training images
• K-d tree algorithm
• Best-bin-first search method

24
Hough Transform

• Cluster reliable model hypotheses
• Search for keys that agree upon a particular
  model pose
• Location, orientation, and scale

25
Affine Transformation