![](bg7.png)
Table 2: Repeatability rate and number of correspondences
for each sequence (overlap error of 40%).
3rd image Sequence
(average)
MSER f-MSER MSER f-MSER
Grafitti 56%(310) 48%(538) 48%(244) 39%(402)
Wall 53%(2093) 57%(2484) 46% (1595) 46% (1786)
Boat 48%(658) 56%(861) 41%(388) 39%(556)
Bark 52%(276) 48%(355) 60% (268) 39%(295)
Bikes 47%(505) 58%(1328) 42%(360) 46%(1002)
Trees 38% (1767) 36% (1796) 33%(1295) 30%(1401)
Leuven 57%(668) 58%(901) 57% (488) 54%(768)
UBC 50%(1114) 63%(1647) 42% (825) 51%(1262)
overlap error of 40%. By analyzing the table, one can
conclude that feature-driven MSERs yield a consid-
erably higher number of correspondences. The lower
average repeatability rate of f-MSERs as seen in some
sequences is mainly due to the decrease of the re-
peatability rate in the last images of the sequence. As
theory confirms, in well-structured scenes, the new
MSERs can yield an increased repeatability. With re-
gard to sequences with viewpoint changes, f-MSERs
have shown a satisfactory repeatability. Nonetheless,
this repeatability can be improved if we build the do-
main under an affine scale-space.
4.2 Completeness Evaluation
How much of the image information is preserved
by local features? Feature sets should capture the
most relevant image content in order to provide
a trustworthy summarized description of the im-
age. This requirement is known in the literature
as completeness. A measure of completeness has
been proposed by Dickscheid et al. (2010). Suc-
cinctly, the incompleteness of a detection corresponds
to the following Hellinger distance: d(p
H
, p
c
) =
q
1
2
∑
~x∈D
(
p
p
H
(~x) −
p
p
c
(~x))
2
, where p
H
corre-
sponds to an entropy density, computed from local
image statistics and p
c
denotes a feature coding den-
sity, inferred from the set of features. We emphasize
that the completeness measure is not a simple way
of evaluating the image coverage of feature sets; it
penalizes local feature sets that contain less informa-
tive patterns despite of their coverage. For the pur-
pose of the evaluation, we defined a dataset compris-
ing images from the Oxford dataset sequences. Each
sequence was represented by its third image. The pa-
rameter settings for the detection coincide with the
ones used in the previous subsection. Table 3 outlines
the results of the completeness evaluation givenby the
Hellinger distance between the two aforementioned
distributions, using feature coding densities computed
from MSER and f-MSER feature sets as well as a
combination of both sets. The analysis of the latter
feature set results allows us to infer on the comple-
Table 3: Dissimilarity measure d(p
H
, p
c
) and number of
regions for the third image of each sequence.
MSER
MSER f-MSER f-MSER+ f-MSER- ∪
f-MSER
Graffiti 0.38(1085) 0.27(2216) 0.33(1130) 0.36(1086) 0.26
Wall 0.2(4433) 0.15(4981) 0.24(2711) 0.26(2270) 0.14
Boat 0.35(2319) 0.28(2720) 0.36 (1393) 0.35(1327) 0.27
Bark 0.29(1940) 0.19(2802) 0.27(1553) 0.31(1249) 0.18
Bikes 0.48(1123) 0.32(2388) 0.4(1311) 0.43(1077) 0.32
Trees 0.24(4821) 0.19(5299) 0.29(2529) 0.26(2770) 0.18
Leuven 0.45(1017) 0.36(1597) 0.44(869) 0.46(728) 0.35
UBC 0.29(2431) 0.26(2619) 0.35(1400) 0.35(1219) 0.24
mentarity of MSER and f-MSER features. Addition-
ally, we have analyzed the completeness of f-MSER+
and f-MSER- features. These results are summarized
in the same table.
It is important to note that MSER-like features are
not among the most complete ones (Dickscheid et al.,
2010). It is readily seen that homogeneous regions
are easily entitled to be classified as extremal regions,
which, in most of the cases, means excluding the most
informative content. Furthermore, the number of fea-
tures detected by the MSER algorithm tends to be
lower than the ones given by other prominent detec-
tors, such as the Hessian-Affine or the Harris-Affine.
One important conclusion to be drawn from Ta-
ble 3 is that f-MSER detection will provide us a
more complete set of features than the one comprised
of standard MSERs. The incompleteness values for
f-MSERs range from 0.15 (“Wall” image) to 0.36
(“Leuven” image), which reflects the high level of
completeness of these features. MSER features are
less complete and, in some cases, even less complete
than f-MSER+ or f-MSER- feature sets. The combi-
nation of MSER and f-MSERs features gives us the
most complete feature sets for each sequence. How-
ever, the complementary between both feature sets
is practically non-existent; the incompleteness values
for MSER ∪ f-MSER features are comparable to the
ones for feature-driven MSERs sets. This result helps
us to conclude that the relevant content preserved by
standard MSERs is also preserved by f-MSERs.
5 CONCLUSIONS AND FUTURE
RESEARCH DIRECTIONS
We have addressed the shortcomings of MSER detec-
tion as well as the desired properties of an image in
order to provide a reliable MSER detection. As result,
we have introduced an alternative domain for MSER
detection. This domain is mainly characterized by the
highlighting of certain boundary-related features and
the simultaneous presence of smooth transitions at the
boundaries. The detection of MSERs on this domain
responds to a higher number of regions than the one
VISAPP 2012 - International Conference on Computer Vision Theory and Applications
496