![](bg2.png)
of distintiveness is largely compensated by the high
efficiency required by online back-end processes for
classification or place recognition which benefit from
a more compact, fast to compare descriptor.
In order to find such compact descriptor, we study
different color spaces and radiometric features, look-
ing for invariance to illumination, point of view and
partial occlusion (section 2). We conclude that a color
descriptor based on the patch dominant color in nor-
malized RGB space provides the best balance between
distinctiveness and compactness. In section 3, this de-
scriptor is compared with the hue based histogram re-
ported in (Pathak et al., 2012), which was previously
proposed for a similar problem, showing that though
the distinctiveness of both are simmilar, our alterna-
tive is advantageous in terms of computation time.
We provide experimental results for place recogni-
tion with plane-based maps using the color descriptor
proposed in this work. We demonstrate that the effi-
ciency of the previous geometry-based solution is sig-
nificantly improved (around 6 times faster) by using
color information. In all the experiments, we compare
our results with with the hue-based histogram.
2 SELECTING A COMPACT
COLOR DESCRIPTOR FOR
PLANAR PATCHES
In this section we address the problem of finding the
simplest color descriptor for a planar patch focused on
the problem of patch matching. This descriptor must
be highly invariant to viewpoint, lighting conditions
and partial occlusion, and also, it must be efficiently
calculated. Note that the utility of this descriptor is
not to unequivocally identify planar patches, but to
prune the search space by adding a very compact ra-
diometric information to the geometric features of the
planar model.
In the context of matching planar patches, a com-
mon solution is that of maximizing the photoconsis-
tency between them (Argiles et al., 2011). The main
limitation of this strategy, which comes as a con-
sequence of the lack of compactness and invariance
of the descriptor, is that maximizing the photocon-
sistency is prohibitively expensive for many applica-
tions, especially when there is not a good initial es-
timation of the registration (e.g. loop closure detec-
tion). Closer to our work are those that describe the
patch radiometric information through its histogram
(Hafner et al., 1995), (Swain and Ballard, 1991). In
this line, (Pathak et al., 2012) posed recently the prob-
lem that we address in this paper: showing how color
information can be exploited to increase the efficiency
of 3D scan registration. A well illuminated scene is
assumed in that work, where the authors adopt a hue
based histogram with 2 extra bins to keep intensity
saturated values (black and white), and test different
measures for histogram distance. However, they do
not take into account the fact that many planar patches
have a single color, so that the histogram contains
redundant information. Also, this descriptor is not
robust to partial occlusion, which is rather common
when doing exploration and mapping.
In this paper, we propose to describe the patch
with its dominant color. A similar strategy is used
in video compression (Manjunath et al., 2001) to de-
fine blobs having the same color. In this way the de-
scriptor storage and the computation of distances are
reduced to a minimum. This is important in a number
of problems where many match combinations have
to be checked in real-time. In order to select such
a descriptor we need to address some issues: first,
the selection of the color space which offers the best
suitability to obtain an invariant and distinctive domi-
nant color (subsection 2.1); second, to define the way
this dominant color is extracted (subsection 2.2); and
third, to adapt the descriptor for cases where the dom-
inant color is not reliable enough (subsection 2.3).
2.1 Selection of the Color Space
In order to obtain a distinctive dominant color, the his-
tograms of the patches must be invariant to illumina-
tion conditions, shading and viewpoint. These charac-
teristics are highly dependent on the color space used
to represent the radiometric information, as we show
in the analysis below. Note also that the fact of se-
lecting the dominant color makes the descriptor in-
herently robust to partial occlusion when the physical
plane has a clearly defined dominant color, which is
the most common situation. If this is not the case,
e.g. a textured plane with different colors, the domi-
nant color is not a good descriptor and it should not
be used for matching.
Different color spaces have been studied in the
context of object recognition in (Gevers and Smeul-
ders, 1999). This work concludes that normal-
ized RGB (rgb), saturation and hue (HS), and the
color models c
1
c
2
c
3
and l
1
l
2
l
3
are highly invariant to
changes in viewing direction and illumination (see ta-
ble 1 for the formulation of these color spaces). Be-
low, we analyze these color spaces for a dataset con-
taining 1000 observations of plane surfaces from dif-
ferent scenarios, spanning diverse viewing conditions
(changing viewpoint and illumination, partial occlu-
sion, etc.). Below we study some relevant properties
ACompactPlanar-patchDescriptorbasedonColor
297