Abstract :

Vanishing points in digital images result from the projection by a pin-hole camera of a set of parallel lines in 3D. Most of the proposed computational methods for detection of vanishing points are forced to rely heavily on additional properties (like orthogonality or coplanarity and equal distance) of the underlying 3D lines in order to avoid spurious responses.

In this work we show that in many images of man-made environments vanishing points alone (without any additional assumption on the underlying 3D lines beyond pure parallelism) are meaningful Gestalts, in the sense of Helmholtz's perception principle, recently proposed for computer vision in [1,3]. This leads to a vanishing point detector with a low false alarms rate and a high precision level, which doesn't rely on any a priory information on the image or calibration parameters, and doesn't require any parameter tuning. The key differences of this detector with respect to previous approaches are the following:

(i) The line segments are themselves detected with almost-zero false alarms rate, by a refinement of the method presented in [2], also based on Helmholtz's principle;

(ii) The threshold to determine a meaningful vanishing point from a large vote in the generalized Hough transform is computed in such a way that it guarantees a low false alarms rate;

(iii) Finally a Minimum Description Length (MDL) criterion is used to further restrict the number of spurious vanishing points.

References:

[1] A. Desolneux, L. Moisan and J-M. Morel, "Edge Detection by Helmholtz Principle", Journal of Mathematical Imaging and Vision, vol. 14, nro. 3, pp 271-284, 2001.

[2] A. Desolneux, L. Moisan and J-M. Morel, " Meaningful Alignments", International Journal of Computer Vision, vol. 40, nro. 1, pp. 7-23, 2000.

[3] A. Desolneux, L. Moisan and J-M. Morel, "Partial Gestalts", submitted, 2001.