Why is the task important?

• Only line correspondences without any point correspondence
• Perspective projection is often present.
• Many images are often available -> a technique treating all data uniformly is needed.

Our solution

• It can be used on both wide base-line stereo and sequences.
• It is based on a factorization method -> a linear solution is computed quickly; subsequent non-linear bundle adjustment is optional.

• Notation: ... scale factor, where	, is the i-th point camera projection matrix
  	is a 6-vector of Pluecker line coordinates lying on the Klein quadric

  projection of space lines into image lines:
  

Our algorithm

1. Estimate from partial reconstructions on triplets of views.
2. Factorize into by the singular value decomposition.
3. Find a transformation bringing rows of and columns of close to the Klein quadric,
4. Project rows of and columns of onto the Klein quadric.

Results

	Cubes	House (Oxford)
	5 images of 14 lines	6 images of 31 lines
	[576x768]	[576x768]
	manual detection	automatic detection
Mean / maximal errors [pxl]
linear method	2.42 / 13.41	0.80 / 10.53
lin. method + bundle adj.	0.47 / 2.12	0.16 / 0.71

References

Line Reconstruction from Many Perspective Images by Factorization. Daniel Martinec and Tomas Pajdla. Accepted at the Computer Vision and Pattern Recognition conference (CVPR), Madison, USA, 2003.

Structure from many perspective images with occlusions. Daniel Martinec and Tomas Pajdla. In Proceedings of the European Conference on Computer Vision (ECCV), pp. 355-369, Springer-Verlag, May 2002. [pdf], [poster]

Demo on reconstruction from points