Maximum-Likelihood Image Matching

1
Maximum-Likelihood Image Matching

• Zheng Lu

2
Introduction

• SSD(sum of squared difference)
• Is not so robust
• A new image matching measure
• Based on maximum-likelihood estimation of
  position
• More robust

3
Maximum-Likelihood Matching

• Set of template feature
• Set of image feature
• The position of template in the image
• t a random variable

4
Maximum-Likelihood Matching

• Distance from each template pixel to the closest
  image pixel.
• Probability density function(PDF) for the
  distance
• Find the t that can maximize this function

5
Estimating the PDF

• The density can be modeled by inliers and
  outliers

6
Estimating the PDF

• The second term should also decrease as d
  increases
• In practice, expected probability density for a
  random outlier is excellent

7
Search Strategy
8

• multi-resolution technique
• divides the space of model positions into cells
  and determines which cells could contain a
  position satisfying the criterion
• Can find the best location, If a conservative
  test is used

9

• c is the center of cell
• distance between the location to template
  edge pixel template mapped by c and any other
  pose in the cell.

10

• The criterion will be