3D surface registration in dental application

1
3D surface registration in dental application

• Gaoyu Xiao
• Jan 11, 2008

2
Content

• Background
• Traditional approach
• Our approach
• Experimental results
• Conclusion

3
Background (1 of 3)
Given the 3D surfaces of a complete tooth and
dental cast model of a patient, the aim is to
estimate the position of the hidden tooth root in
the dental cast model.
4
Background (2 of 3)

• Acquisition of 3D surfaces using Cyberware 3D
  laser scanner

5
Background (3 of 3)

• Crown surface extraction from dental cast
  (Kondo,TMI04)

6
Traditional approach (1 of 3)

• Use certain surface descriptor, e.g. spin-image
  (Johnson,99), to find the point correspondences,
  based on which 3D spatial transformation can be
  computed.
• No need for initial estimate or feature
  extraction.
• False point correspondences may occur, which need
  to be removed before the computation of the 3D
  transformation.

7
Traditional approach (2 of 3)

• Spin-image (Johnson,PAMI99)

8
Traditional approach (3 of 3)

• Use geometric consistency measure to remove false
  point correspondences.

where C1s1,m1 and C2s2,m2 are two pairs of
point correspondences.

• Con Every two correspondences need to be
  compared with each other.

9
Our approach (1 of 4)

• Use spin-image to find the candidate point
  correspondences
• Compute spin-images only for candidate vertices.

10
Our approach (2 of 4)

• False point correspondences after comparison of
  spin-images

11
Our approach (3 of 4)

• Efficient rejection of false point
  correspondences
• Use 4-point tuples to compute the corresponding
  centroid (i.e. translation).

12
Our approach (4 of 4)

• Efficient rejection of false point
  correspondences
• Use modified-ISODATA algorithm to cluster the
  transformation parameters
• Further iterative refinement (Umeyama,PAMI91) of
  transformation parameters.

13
Experimental results (1 of 4)

• Teeth without outliers

14
Experimental results (2 of 4)

• Teeth with outliers added

15
Experimental results (3 of 4)

• Numerical comparison

16
Experimental results (4 of 4)

• Bunny

17
Conclusion

• Our approach is especially applicable to the 3D
  partial registration of smooth surface objects
  that may also have many regionally similar
  surface patches.

18
Thank You