CS698 Project Biologically Inspired

1
CS698 ProjectBiologically Inspired

• This presentation will probably involve audience
  discussion, which will create action items. Use
  PowerPoint to keep track of these action items
  during your presentation
• In Slide Show, click on the right mouse button
• Select Meeting Minder
• Select the Action Items tab
• Type in action items as they come up
• Click OK to dismiss this box
• This will automatically create an Action Item
  slide at the end of your presentation with your
  points entered.

• Object Recognition Through SIFT Recognition Keys

2
Motivation

• Vision sense modality conveying huge
  information
• Object Recognition Tough task
• Biological vision recognize objects under
  different
• Orientations
• illuminations
• and  at different scales

3
SIFT key algorithm by Lowe

• SIFT stands for scale invariant feature transform
• Provides features for object recognition
• Patented by university of British Columbia
• Similar to the one used in primate visual system

4
Other work/approaches

• Object segmentation and correlation maximization
• Eigenspace matching
• Isolated objects / pre-segmented image
• Schmid Mohr Harris Corner detector to identify
  interest points, create a local image decriptor
  vector i.e. independent of orientation
• Examines at a single scale

5
Steps

• Scale space peak selection
• Potential locations for finding features
• Key point Localization
• Accurately locating the feature key
• Orientation Assignment
• Assigning orientation to the keys
• Key point descriptor
• Describing the key as a high dimensional vector

6
Scale space peak selection

• Scale space of an image is defined as
• L(x,y,sig)G(x,y,sig)I(x,y)
• Locating stable key points by repeatedly
  smoothing and downsampling an input image and
  subtracting adjacent levels to create a pyramid
  of difference-of-Gaussian image.
• D(x,y,sig) L(x,y,ksig) L(x,y,sig)

7
Scale space pyramid
8
Peak detection
9
Key point localization

• To the first approximation the location of key is
  the place where peak is detected
• Better Fit a 3D paraboloid through the
  neighboring points to locate the maxima in the
  scale space
• Reject those points with a low contrast value
• Eliminate poor edge locations

10
(No Transcript)
11
Orientation assignment Finding gradient
magnitude m and orientation theta
12
Orientation histogram

• Orientation histogram formed from the gradient
  orientations at all sample points within a
  circular window around the key point.
• Each sample added to the histogram is weighted by
  its gradient magnitude and by Gaussian with sigma
  3 times that of the current scale.
• 36 bins covering 360 degrees

13
Assigning orientation

• Locate peaks in the histogram
• Highest peak is detected and those lying within
  80 of it are also detected and used to create
  different keys at the same location.
• 15 of the points are assigned multiple
  orientation
• Contribute significantly to the stability of
  matching

14
Local Image descriptor

• Again orientation histograms are formed (with
  relative orientation)
• 4X4 region around the key point
• Magnitude weighted with gaussian
• 8X4X4128 element feature vector normalized to
  reduce effect of illumination change
• No gradient magnitude is allowed to be greater
  than 0.2
• Threshold at 0.2
• renormalized

15
Key point descriptor
16
Similarity to IT cortex

• Complex neurons respond to a gradient at a
  particular orientation.
• Location of the feature can shift over a small
  receptive field.
• Edelman, Intrator, and Poggio (1997) hypothesized
  that the function of the cells is allow for
  matching and recognition of 3D objects from a
  range of view points.
• Experiments show better recognition accuracy for
  3D objects rotated in depth by up to 20 degrees

17
Key point matching

• Match the key points against a database of that
  obtained from training images.
• Find the nearest neighbor i.e. a keypoint with
  minimum Euclidean distance.
• Efficient Nearest Neighbor matching
• Best bin first algorithm a modification of k-d
  tree search
• Approx solution
• Second nearest neighbor distance restriction

18
Further Improvements

• Clustering keys together
• Hough transform
• Solution to affine transforms

19
Bibliography

• Distinctive image features from scale-invariant
  keypoints
• David G. Lowe, accepted for publication in the
  International Journal of Computer Vision, 2004.  
  
• Object recognition from local scale-invariant
  features
• David G. Lowe, International Conference on
  Computer Vision, Corfu, Greece (September 1999),
  pp. 1150-1157.
• Local feature view clustering for 3D object
  recognition,
• David G. Lowe, IEEE Conference on Computer Vision
  and Pattern Recognition, Kauai, Hawaii (December
  2001), pp. 682-688.
• Invariant Features from Interest Point Groups,
• Matthew Brown and David G. Lowe British Machine
  Vision Conference, BMVC 2002, Cardiff, Wales
  (September 2002).
• Local grayvalue invariants for image retrievals,
• Schmid, C., and R. Mohr, IEEE PAMI, 19, 5 (1997),
  pp 530-34