Image Registration using Triangular Mesh

1
Image Registration using Triangular Mesh

• Ben Yip, University of Sydney
• Jesse S Jin, University of Newcastle

2
Image registration

• A fundamental problem in computer vision
• Commonly needed in 2 areas
• Motion tracking
• Binocular stereo matching
• There are two schools of image registration
• Find feature points in each image and then match
  up
• Find feature points in one image first and then
  seeks the corresponding points. (Our approach)

3
Feature points

• Commonly by edge detection (or similar)
• Our approach feature points are uniformly
  scattered in the image as the nodes of a
  triangular mesh
• Advantages
• Eliminate errors due to wrong feature points
  finding
• Guarantees an even distribution

4
Mapping a feature point

• Mapping a feature point can be viewed as
  different combinations of choices for (Brown 92)
  
• Feature space. We use 9x9 RGB block.
• Search space. We use a 79x79 RGB block.
• Search strategy. We start from the expected
  position if there are clues.
• Similarity metric. We use average colour
  difference. More precisely, it is a measure of
  dissimilarity.

5
Dissimilarity Metric
In our case, maxX 39, maxY 39, because we use
a 79x79 block
6
Problems with feature mapping

• It is a local process, there is no consideration
  of the nearby mapping results. This may results
  in a crossing. Crossing may be caused by
  occlusion (correct), or it is a mismatch (error).
  
• Our approach targets for a smooth mapping, and
  hence the resulted triangular mesh should not
  have crossings.

7
Our approach

• Sort each of the feature points by their
  similarity scores in ascending order.
• Add each of the sorted feature points into the
  output triangular mesh.
• If the newly added feature point generates a
  crossing, this feature point needs to be
  re-mapped.
• Re-mapping process is the same as the primary
  feature point mapping except that the search
  space is now bounded by the top, left, bottom,
  right neighbors of the feature point.
• This guarantees a smooth mapping

8
Evaluation 1
INPUT
OUTPUT
ACTUAL
9
Evaluation - 2
INPUT
OUTPUT
10
Conclusion

• Our approach is unique in a way that it uses
  triangular mesh with image registration and
  having uniformly distributed feature points.
• It is designed for images of objects with smooth
  depth. It performs well relatively on images
  that do not have distinctive feature points.
• Our approach of using triangular mesh is
  convenient to any applications that need to deal
  with triangular mesh in the later stage.