Widely Separated View Matching
Desirable properties that a good feature point
detector should exhibit:
A new feature point detector is proposed which uses a
wedge model to characterize corners by their orientation and
angular width. This detector is compared to two popular feature
point detectors: the Harris and SUSAN detectors, on the basis of
some defined quality attributes. It is also shown how feature points
between widely separated views can be matched by using
information provided by the detector to approximate local affine
- Accuracy , or the ability to consistently
detect a given image pattern, at the exact same location, in spite
of minor variability in intensity values, in orientation, or in
scale. This property is significant when the detector is tuned to
detect well-defined structures, such as specific types of
junctions or corners. Accuracy can be assessed by measuring the
alignment of extracted features located on a straight line, or by measuring the distance between a detected
junction and the point of intersection of the two lines defining
- Robustness, or insensitivity to noise.
Detection on noisy images can produce false positives,
corresponding to noise patterns rather than true feature points.
Also, in the presence of noise, some feature points can be lost or
not localized properly. Robustness can be evaluated empirically,
- Sensitivity of the detection, that is the ability to detect
feature points in low contrast conditions. Usually, there are some
parameters which control the sensitivity of a detector, and most
often, a tradeoff exists between sensitivity and robustness.
- Stability of the detection. A detected feature point should continue to be
detected after an image undergoes some geometrical transformation
(especially the perspective deformation due to a change in
viewpoint), and under different conditions of illumination. This
is essential in the context of multi-view matching. A good measure
of stability is the repeatability of detected features across
- Controllability of a detector. This is mainly determined by the number of
parameters which control its behavior, and their relative
sensitivity. It is certainly useful to be able to control the
number of feature points which will be selected in an image,
ideally using a single control parameter. Other parameters might
also be used to filter out certain kinds of features, not of
interest in a specific application. However, the effect of each
parameter should be specific and predictable enough to allow an
- Richness of the information provided about the detected feature points.
When a detector returns various characteristics of feature points,
and not just a strength measure, the additional information can be
exploited in the task to follow. For example, it could be used to
classify feature points into categories (e.g. type of junction), or in matching, as a tool to normalize the
patterns being matched.
- Variability in the characteristics of detected feature
points. A high variability ensures that several feature points are
detected, regardless of the nature of the image under analysis.
Good variability is also critical for matching, where the feature
points must be easily distinguishable from each other.
- Complexity of the detector, or the speed at which
it identifies corners in an image. In many applications, feature
point detection is a preprocessing operation that must be
performed at frame rate. However, a comparison of detection speeds
is difficult to achieve, since the efficiency of a given feature
point detector depends on its implementation. Nevertheless, the
complexity of the operations which must be applied to each image
point must remains sufficiently low.
Here (non-optimized version)
Robert Laganière and Étienne Vincent,
Detecting and matching feature points
in Journal of Visual Communication and Image Representation
, vol. 16, issue 1, pp. 38-54, Febuary 2005.
Elsevier Science Direct
Robert Laganière and Étienne Vincent,
Wedge-based Corner Model for Widely Separated Views Matching,
in Proc. 16th International Conference on Pattern Recognition, vol. 3, Quebec City, Canada, August 2002.
PDF [1.2 mb]
Copyright © 2004