Indexing Based On Scale Invariant Interest Points
Indexing Based On Scale Invariant Interest Points
Indexing Based On Scale Invariant Interest Points
H!0,I /S@
where denotes the 8-neighbourhood of the point .
In order to obtain a more compact representation, we ver-
ify for each of the candidate points found on different levels Figure 4: Repeatability of interest point detectors with re-
if it forms a maximum in the scale direction. The Laplacian spect to scale changes.
is used for selection.
H!0MI / H!0,I , ?H!0MI / H!0,I , 4. Robust matching and indexing
H!,I6 </ @)% In the following we briefly describe our robust matching
Figure 5 shows the scale-space representation for two and indexing algorithms. The two algorithms are based on
real images with points detected by the Harris-Laplacian the same initial steps:
method. For these two images of the same object imaged at 1. Extraction of Harris-Laplacian interest points (cf. sec-
different scales we present for each scale level the selected tion 3).
points. There are many point-to-point correspondences be- 2. Computation of a descriptor for each point at its char-
tween the levels for which the scale ratio corresponds to the acteristic scale. Descriptors are invariant to image ro-
real scale change between the images (indicated by point- tation and affine illumination changes. They are robust
ers). Additionally, very few points are detected in the same to small perspective deformations.
1.92 s=1.2 s=2.4 s=4.8 s=9.6
Figure 5: Points detected on different resolution levels with the Harris-Laplacian method.
3. Comparison of descriptors based on the Mahalanobis trix. A model selection algorithm [6] can of course be used
distance. to automatically decide what transformation is the most ap-
Interest points. To extract interest points we have used propriate one.
a scale representation with 17 resolution levels. The initial Indexing. A voting algorithm is used to select the most
scale K is 1.5 and the factor J between two levels of resolu- similar images in the database. This makes retrieval robust
tion is 1.2. The parameter is set to 0.06 and the thresholds to mismatches as well as outliers. For each point of a query
@ and @ % are set to 1500 and 10, respectively. image, its descriptor is compared to the descriptors in the
Descriptors. Our descriptors are Gaussian derivatives database. If the distance is less than a fixed threshold , a vote
which are computed at the characteristic scale. Invariance is added to the corresponding database image. Note that a
to rotation is obtained by “steering” the derivatives in the point cannot vote several times for the same database im-
direction of the gradient [4]. To obtain a stable estimation age. The database image with the highest number of votes
of the gradient direction, we use the peak in a histogram of is the most similar one.
local gradient orientations. Invariance to the affine inten-
5. Experimental results
sity changes is obtained by dividing the derivatives by the
steered first derivative. Using up to 4th order derivatives, In the following, we validate our detection algorithm by
we obtain descriptors of dimension 12. matching and indexing results. Figure 6 illustrates the dif-
ferent steps of our matching algorithm. In this example the
Comparison of descriptors. The similarity of descriptors
two images are taken from the same viewpoint, but with a
is measured by the Mahalanobis distance. This distance
change in focal length and image orientation. The top row
requires the estimation of the covariance matrix which
shows the detected interest points. There are 190 and 213
encapsulates signal noise, variations in photometry, inaccu-
points detected in the left and right images, respectively.
racy of interest point location, and so forth. is estimated
The number of detected points is about equivalent to results
statistically over a large set of image samples.
obtained by a standard interest point detector. This clearly
Robust matching. To robustly match two images, we first shows the selectivity of our point detection method. If no
determine point-to-point correspondences. We select for scale peak selection had been used, more than 2000 points
each descriptor in the first image the most similar descrip- would be detected. The middle row shows the 58 matches
tor in the second image based on the Mahalanobis distance. obtained during the initial matching phase. The bottom row
If the distance is below a threshold the match is kept. This displays the 32 inliers to the estimated homography, all of
allows us to obtain a set of initial matches. A robust esti- which are correct. The estimated scale factor between the
mation of the transformation between the two images based two images is 4.9 and the estimated rotation angle is 19 de-
on RANdom SAmple Consensus (RANSAC) allows to re- grees.
ject inconsistent matches. For our experimental results the Figure 7 shows an example for a 3D scene where the
transformation is either a homography or a fundamental ma- fundamental matrix is used for verification. There are 180
and 176 detected points detected in the left and right im- database (second row) was correctly retrieved, that is it was
ages. The number of initial matches is 23 and there are 14 the most similar one. The approximate scale factor is given
inliers to the robustly estimated fundamental matrix, all of in row three. The changes between the image pairs (first and
them correct. Note that the images are taken from different second row) include important changes in the focal length,
viewpoints, the transformation includes a scale change, an for example 5.8 for the image pair (a). They also include
image rotation as well as a change in the viewing angle. The important changes in viewpoint, for example for pair (b).
building in the middle is almost half occluded. Furthermore, they include important illumination changes
(image pair (e)).
Extracted interest points
Figure 8: The first row shows some of the query images. The second row shows the most similar images in the database, all
of them are correct. The approximative scale factor between query image and database image is given in row three.