SIFT Scale Invariant Feature Transform

1
SIFTScale Invariant Feature Transform

• A method for extracting distinctive invariant
  features

2
Back ground

• Application
• Matching between different views of a object or a
  scene.
• Invariant to
• image scaling
• rotation
• robust matching across substantial range of
• Distortion
• Change in 3D view point
• Addition of noise
• Change in illumination

3
Example
4
Application

• Object recognition
• View matching for 3D reconstruction
• Motion tracking and segmentation
• Robot localization
• Image panorama assembly
• Epipolar calibration

5
Major stages to generate the image features

• Scale space extrema detection
• Key point localization
• Orientation assignment
• Key point descriptor

6
Scale-space extrema detection

• The convolution of the difference of
• different variable-scale Gaussian with the image

7
Scale-space extrema detection

• Compare a pixel with its 26 neighbors in 33
  regions at the current and adjacent scales
• Find the extrema response in the schema of
  response

8
Keypoints localization

• Reject keypoints with low contrast
• Eliminating edges response

9
Orientation assignment

• Compute the gradient magnitude and orientation
• An orientation histogram is formed

10
Local image descriptor
11
properties

• Invariant to
• image scaling
• rotation
• robust matching across substantial range of
• Distortion
• Change in 3D view point
• Addition of noise
• Change in illumination

12
Strong points

• No need predetermined scale
• Robustness for affine change

13
Possible research

• Illumination-invariant color descriptor
• Local texture measures (comparing to the single
  spatial frequency)
• To recognize objects categories by learning
  features

14
The process of object recognition

• Keypoint matching
• Efficient nearest neighbor indexing
• Clustering with the Hough transform

15
Points

• Distinctiveness high dimensional vectors
• robustness in extracting small objects among
  clutter large number of keypoints are extracted
• available for matching small and highly occluded
  objects keypoints are detected over a complete
  range of scales

16
(No Transcript)
17
(No Transcript)
18
(No Transcript)