Projective 3D-reconstruction from Perspective Images
with Occlusions and Outliers

Daniel Martinec
Center for Machine Perception, Czech Technical University in Prague

Why is the task important?

• Perspective projection, occlusions and outliers are 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 has automatical outlier detection.
• It is based on a factorization method -> linear solution is computed fast; subsequent non-linear bundle adjustment is optional.
• Outliers are rejected using a simplified RANSAC technique. Significant speed-up is achieved using reconstructions from partial data.
• Notation: ... outlier, ... occlusion, ... projective depth ()
  

Results

	House	Dinosaur (Oxford)	Temple (Leuven)	Castle (Leuven)	Valbonne (Oxford)
	10 images of 203 points	36 images of 4983 points	5 images of 696 points	22 images of 1822 points	14 images of 376 points
	[2952x2003]	[720x576]	[867x591]	[768x576]	[768x512]
	manual detection	Harris' operator	Harris' operator	Harris' operator	extremal regions
	47.83 % occlusions	90.84 % occlusions	46.32 % occlusions	74.99 % occlusions	73.27 % occlusions
Mean errors [pxl]
With outliers
linear method		1.76	0.49	41.97
lin. method + bundle adj.		0.64	0.23	11.97
Outliers detected and removed			corresp. in > 2 images
linear method	3.91	0.57	0.37	0.65	0.50
lin. method + bundle adj.	1.44	0.39	0.27	0.22	0.45

References

Presentation given at the Pattern Recognition and Computer Vision Colloquium, Summer 2002 [pdf]