IMPROVING JUNCTION DETECTION BY SEMANTIC

INTERPRETATION

Sinan Kalkan

Bernstein Centre for Computational Neuroscience, Univ. of G

ottingen, Germany

Shi Yan, Florian Pilz

Medialogy Lab, Aalborg Univ. Copenhagen, Denmark

Norbert Kr

uger

Cognitive Vision Group, Univ. of Southern Denmark, Denmark

Keywords:

Junction Detection, Junction Positioning, Junction Representation.

Abstract:

Every junction detector has a set of thresholds to make decisions about the junctionness of image points.

Low-contrast junctions may pass such thresholds and may not be detected. Lowering the thresholds to ﬁnd

such junctions will lead to spurious junction detections at other image points. In this paper, we implement a

junction-regularity measure to improve localization of junctions, and we develop a method to create semantic

interpretations of arbitrary junction conﬁgurations at improved junction positions. We propose to utilize such

a semantic interpretation as a feedback mechanism to ﬁlter false-positive junctions. We show results of our

proposals on natural images using Harris and SUSAN operators as well as a continuous concept of intrinsic

dimensionality.

1 INTRODUCTION

Junctions are utilized in computer vision and image

processing for tasks that especially require ﬁnding

correspondences between different views of the same

scene, mainly due to their distinctiveness, seldomness

and stability.

Correct localization of junctions

is crucial be-

cause even small errors in localization lead to wrong

interpretations of the scene (Rohr, 1992). Neverthe-

less, it is shown in (Deriche and Giraudon, 1993;

Rohr, 1992) that energy-based junction detection

methods smooth out junctions and face the problem

of wrong localization.

Junctions also have the property of being inter-

pretable: i.e., you can construct a meaningful inter-

pretation about how the junction is formed, as pro-

posed in (Parida et al., 1998; Rohr, 1992). Such a se-

mantic interpretation (SI) can be utilized in rigid body

motion estimation, depth estimation, feature match-

ing etc. and should be more robust than a single junc-

In this paper, corners are considered to be a special case

of junctions, and the term ’corner’ is avoided.

tionness measure in identiﬁcation of junctions and in

correspondence ﬁnding.

Junction detectors, no matter what the underlying

methods are, have to make a decision about the junc-

tionness of image points. The decision is made by a

set of automatically or manually set thresholds (on a

set of measures) that determine the sensitivity of the

algorithm to contrast (in most of the cases, a high

threshold means low sensitivity and vice versa). On

the other hand, a method that utilizes a junction de-

tector requires the detector to be complete: i.e., the

detector should be able to detect all the junctions that

represent the image.

The relation between sensitivity and completeness

presumably looks like as plotted in ﬁgure 1(a). In-

creasing the sensitivity increases not only the com-

pleteness of a detector

but also increases the amount

of false-positives, or ’spuriousness’, of the detector as

illustrated in ﬁgure 1(b). These observations suggest

that spuriousness and completeness are two compet-

Exact shape of this relation might be different in real

world; however, the authors claim that completeness should

be still an increasing function of sensitivity in any case.

264

Kalkan S., Yan S., Pilz F. and Krüger N. (2007).

IMPROVING JUNCTION DETECTION BY SEMANTIC INTERPRETATION.

In Proceedings of the Second International Conference on Computer Vision Theory and Applications - IFP/IA, pages 264-271

 SciTePress