Content uploaded by Marc Pollefeys
Author content
All content in this area was uploaded by Marc Pollefeys on Nov 25, 2015
Content may be subject to copyright.
Noname manuscript No.
(will be inserted by the editor)
Image Based Geo-Localization in the Alps
Olivier Saurer ·Georges Baatz ·Kevin Köser ·
L’ubor Ladický ·Marc Pollefeys
Received: date / Accepted: date
Abstract Given a picture taken somewhere in the world, automatic geo-localization
of such an image is an extremely useful task especially for historical and forensic
sciences, documentation purposes, organization of the world’s photographs and in-
telligence applications. While tremendous progress has been made over the last years
in visual location recognition within a single city, localization in natural environ-
ments is much more difficult, since vegetation, illumination, seasonal changes make
appearance-only approaches impractical. In this work, we target mountainous terrain
and use digital elevation models to extract representations for fast visual database
lookup. We propose an automated approach for very large scale visual localization
that can efficiently exploit visual information (contours) and geometric constraints
(consistent orientation) at the same time. We validate the system at the scale of
Switzerland (40000km2) using over 1000 landscape query images with ground truth
GPS position.
Keywords Geo-Localization ·Localization ·Camera Calibration ·Computer Vision
1 Introduction and Previous Work
In intelligence and forensic scenarios as well as for searching archives and organis-
ing photo collections, automatic image-based location recognition is a challenging
Olivier Saurer
ETH Zürich, Switzerland, E-mail: saurero@inf.ethz.ch
Georges Baatz
Google Inc., Zürich, Switzerland, E-mail: gbaatz@google.com
Kevin Köser
GEOMAR Helmholtz Centre for Ocean Research Kiel, Germany, E-mail: kkoeser@geomar.de
L’ubor Ladický
ETH Zürich, Switzerland, E-mail: lubor.ladicky@inf.ethz.ch
Marc Pollefeys
ETH Zürich, Switzerland, E-mail: marc.pollefeys@inf.ethz.ch
2 Olivier Saurer et al.
task that would be extremely useful when solved. In such applications GPS tags are
typically not available in the images requiring a fully image-based approach for geo-
localization. Over the last years progress has been made in urban scenarios, in par-
ticular with stable man-made structures that persist over time. However, recognizing
the camera location in natural environments is substantially more challenging, since
vegetation changes rapidly during seasons, and lighting and weather conditions (e.g.
snow lines) make the use of appearance-based techniques (e.g., patch-based local im-
age features [28,8]) very difficult. Additionally, dense street-level imagery is limited
to cities and major roads, and for mountains or for the countryside only aerial footage
exists, which is much harder to relate with terrestrial imagery.
In this work we give a more in depth discussion on camera geo-localization in
natural environments. In particular we focus on recognizing the skyline in a query
image, given a digital elevation model (DEM) of a country — or ultimately, the world.
In contrast to previous work of matching e.g. a peak in the image to a set of mountains
known to be nearby, we aggregate shape information across the whole skyline (not
only the peaks) and search for a similar configuration of basic shapes in a large scale
database that is organized to allow for query images of largely different fields of view.
The method is based on sky segmentation, either automatic or easily supported by an
operator for challenging pictures such as those with reflection, occlusion or taken
from inside a cable car.
Contributions.
A preliminary version of this system was presented in [2]. This work provides a more
detailed analysis and evaluation of the system and improves upon the skyline seg-
mentation. The main contributions are a novel method for robust contour encoding as
well as two different voting schemes to solve the large scale camera pose recognition
from contours. The first scheme operates only in descriptor space (it verifies where
in the model a panoramic skyline is most likely to contain the current query picture)
while the second one is a combined vote in descriptor and rotation space. We validate
the whole approach using a public digital elevation model of Switzerland that covers
more than 40000km2and a set of over 1000 images with ground truth GPS position.
In particular we show the improvements of all novel contributions compared to a
baseline implementation motivated by classical bag-of-words [31] based techniques
like [8]. In addition we proposed a semi-automatic skyline segmentation technique,
based on a dynamic programming approach. Furthermore, we demonstrate that the
skyline is highly informative and can be used effectively for localization.
Previous Work.
To the best of our knowledge this is the first attempt to localize photographs of
natural environments at large scale based on a digital elevation model. The closest
works to ours are smaller scale navigation and localization in robotics [37,32], and
building/location recognition in cities [28,1,8, 26, 34, 4] or with respect to commu-
nity photo collections of popular landmarks [19]. These, however, do not apply to
landscape scenes of changing weather, vegetation, snowlines, or lighting conditions.
Image Based Geo-Localization in the Alps 3
(a) (b) (c) (d) (e)
Fig. 1 Different stages in the proposed pipeline: (a) Query image somewhere in Switzerland, (b) sky
segmentation, (c) sample set of extracted 10◦contourlets, (d) recognized geo-location in digital elevation
model, (e) overlaid skyline at retrieved position.
The robotics community has considered the problem of robot navigation and robot
localization using digital elevation models for quite some time. Talluri et al. [33]
reason about intersection of known viewing ray directions (north, east, south, west)
with the skyline and relies thus on the availability of 360◦panoramic query contours
and the knowledge of vehicle orientation (i.e. north direction). Thompson et al. [35]
suggest general concepts of how to estimate pose and propose a hypothesize and
verification scheme. They also rely on known view orientation and match viewpoint-
independent features (peaks, saddle points, etc.) of a DEM to features found in the
query image, ignoring most of the signal encoded in the skyline. In [11], computer vi-
sion techniques are used to extract mountain peaks which are matched to a database
of nearby mountains to support a remote operator in navigation. However, we be-
lieve that their approach of considering relative positions of absolute peaks detected
in a DEM is too restrictive and would not scale to our orders of magnitude larger
problem, in particular with respect to less discriminative locations. Naval et al. [24]
proposes to first match three features of a contour to a DEM and estimate an ini-
tial pose from that before doing a non-linear refinement. Also here the initial step of
finding three correct correspondences is a challenging task in a larger scale database.
Stein et al. [32] assumes panoramic query data with known heading, and computes
super-segments on a polygon fit, however descriptiveness/robustness is not evalu-
ated on a bigger scale, while [10] introduces a probabilistic formulation for a similar
setting. The key point is that going from tens of potential locations to millions of
locations requires a conceptually different approach, since exhaustive image compar-
ison or trying all possible “mountain peaks” simply does not scale up to a large-scale
geo-localization problems. Similarly, for urban localization, in [27] an upward look-
ing 180◦field-of-view fisheye is used for navigation in urban canyons. They render
untextured city models near the predicted pose and extract contours for compari-
son with the query image. A similar approach was recently proposed by Taneja et al.
[34], where panoramic images are aligned to a cadastral 3D model by maximizing the
overlap between the panoramic image and the rendered model. In [26] Ramalingam
et al. propose a general framework to solve for the camera pose using 3D-to-2D point
and line correspondences between the 3D model and the query image. The approach
requires an initial correspondence match, which is propagated to the next image using
appearance based matching techniques. These approaches are meant as local meth-
ods for navigation or pose refinement. Also recently, in [3] Baboud et al. optimize
the camera orientation given the exact position, i.e. they estimate the viewing direc-
tion given a good GPS tag. In [4] Bansal et al. propose a novel correspondence-free
geo-localization approach in urban environments. They match corners and roof-line
4 Olivier Saurer et al.
edges of buildings to a database of 3D corners and direction vectors previously ex-
tracted from a DEM. None of the above mentioned systems considered recognition
and localization in natural environments at large scale.
On the earth scale, Hays et al. [13] source photo collections and aim at learning
location probability based on color, texture, and other image-based statistics. Con-
ceptually, this is not meant to find an exact pose based on geometric considerations
but rather discriminates landscapes or cities with different (appearance) characteris-
tics on a global scale. In [18] Lalonde et al. exploit the position of the sun (given the
time) for geo-localization. In the same work it is also shown that identifying a large
piece of clear sky without haze provides information about the camera pose (although
impressive given the data, over 100km mean localization error is reported). Both ap-
proaches are appealing for excluding large parts of the earth from further search but
do not aim at exactly localizing the camera within a few hundred meters.
Besides attacking the DEM-based, large scale geo-localization problem we pro-
pose new techniques that might also be transferred to bag-of-words approaches based
on local image patches (e.g. [31,28,8]). Those approaches typically rely on pure
occurrence-based statistics (visual word histogram) to generate a first list of hypothe-
ses and only for the top candidates geometric consistency of matches is verified.
Such a strategy fails in cases where pure feature coocurrence is not discriminative
but where the relative locations of the features are important. Here, we propose to do
a (weak) geometric verification already in the histogram distance phase. Furthermore,
we show also a representation that tolerates largely different document sizes (allow-
ing to compare a panorama in the database to an image with an order of magnitude
smaller field-of-view).
2 Mountain Recognition Approach
The location recognition problem in its general form is a six-dimensional problem,
since three position and three orientation parameters need to be estimated. We make
the assumption that the photographs are taken not too far off the ground and use the
fact that people rarely twist the camera relative to the horizon [7] (e.g. small roll).
We propose a method to solve that problem using the outlines of mountains against
the sky (i.e. the skyline). For the visual database we seek a representation that is
robust with respect to tilt of the camera which means that we are effectively left with
estimating the 2D position (latitude and longitude) on the digital elevation model
and the viewing direction of the camera. The visible skyline of the DEM is extracted
offline at regular grid positions (360◦at each position) and represented by a collection
of vector-quantized local contourlets (contour words, similar in spirit to visual words
obtained from quantized image patch descriptors [31]). In contrast to visual word
based approaches, additionally an individual viewing angle αd(αd∈[0;2π]) relative
to north direction is stored. At query time, a skyline segmentation technique is applied
that copes with the often present haze and also allows for user interaction in case of
incorrect segmentation. Subsequently the extracted contour is robustly described by
a set of local contourlets plus their relative angular distance αqwith respect to the
optical axis of the camera. The contour words are represented as an inverted file
Image Based Geo-Localization in the Alps 5
system, which is used to query the most promising location. At the same time the
inverted file also votes for the viewing direction, which is a geometric verification
integrated in the bag-of-words search.
2.1 Processing the Query Image
2.1.1 Sky Segmentation
The estimation of the visible skyline can be cast as a foreground-background segmen-
tation problem. As we assume almost no camera roll and since overhanging structures
are not modelled by the 2.5D DEM, finding the highest foreground pixel (foreground
height) for each image column provides an good approximation and allows for a
dynamic programming solution, as proposed in [20,5]. To obtain the data term for
a candidate height in a column we sum all foreground costs below the candidate
contour and all sky costs above the contour. The assumption is, when traversing the
skyline, there should be a local evidence in terms of an orthogonal gradient (similar
in spirit to flux maximization [36] or contrast sensitive smoothness assumptions [6,
15] in general 2D segmentation).
We express the segmentation problem in terms of an energy:
E=
width
∑
x=1
Ed(x) + λ
width−1
∑
x=1
Es(x,x+1),(1)
where Edrepresents the data term, Esthe smoothness term and λis a weighting
factor. The data term Ed(x)in one column xevaluates the cost of all pixel below it
to be assigned a foreground label while all pixels above it are assigned a background
(sky) label. The cost is incorporated into the optimization framework as a standard
negative-log-likelihood:
Ed=
k−1
∑
i=1
−logh(F|zi) +
height
∑
i=k
−logh(B|zi),(2)
where h(F|zi)denotes the probability of pixel zibeing assigned to the foreground
Fmodel and h(B|zi)the probability of a pixel being assigned to the background B
model. The likelihoods h(z|F)and h(z|B)are computed by the pixel-wise classifier,
jointly trained using contextual and superpixel based feature representations [17].
The contextual part of the feature vector [30,16] consists of a concatenation of
bag-of-words representations over a fixed random set of 200 rectangles, placed rel-
ative to the corresponding pixel. These bag-of-words representations are built using
4 dense features - textons [22], local ternary patterns [14], self-similarity [29] and
dense SIFT [21], each one quantized to 512 clusters using standard K-means clus-
tering. For each pixel the superpixel part of the feature vector is the concatenation
of a bag-of-words representations of a corresponding superpixel [17] from each un-
supervised segmentation. Four superpixel segmentations are obtained by varying the
parameters of the MeanShift algorithm [9], see Fig. 2. Pixels, belonging to the same
6 Olivier Saurer et al.
(a) (b) (c) (d)
Fig. 2 Superpixel based segmentation: (a) Input image. (b) MeanShift filtered image. (c) MeanShift region
boundaries. (d) Final segmentation.
segment, share a large part of the feature vector, and thus tend to have the same labels,
leading to segmentations, that follow semantic boundaries.
The most discriminative weak features are found using AdaBoost [12]. The con-
textual feature representations are evaluated on the fly using integral images [30],
the superpixel part is evaluated once and kept in memory. The classifier is trained
independently for 5 colour spaces - Lab, Luv, Grey, Opponent and Rgb. The final
likelihood is calculated as an average of these 5 classfiers.
The pairwise smoothness term is formulated as:
Es(x,x+1) = ∑
i∈C
exp−d⊤Rgi
λ||d|| ,(3)
where Cis the set of pixels connecting pixel znin column xand zmin column
x+1 along the Manhattan path (path along the horizontal and vertical direction), d
is the direct connection vector between znand zm,giis the image gradient at pixel
i,Rrepresents a 90 degree rotation matrix and λis set to the mean of d⊤Rgifor
each image. The intuition is, that all pixels on the contour should have a gradient
orthogonal to the skyline.
Given the energy terms defined in Eq. (2) and (3), the segmentation is obtained by
minimizing Eq. (1) using dynamic programming. Our framework also allows for user
interaction, where simple strokes can mark foreground or background (sky) in the
query image. In case of a foreground labelling this forces all pixel below the stroke
to be labels as foreground and in case of a backround stroke, the stroke pixel and all
pixels above it are marked as background (sky). This provides a simple and effective
means to correct for very challenging situations, where buildings and trees partially
occlude the skyline.
2.1.2 Contourlet Extraction
In the field of shape recognition, there are many shape description techniques that
deal with closed contours, e.g. [23]. However, recognition based on partial contours
is still a largely unsolved problem, because it is difficult to find representations in-
variant to viewpoint. For the sake of robustness to occlusion, to noise and systematic
errors (inaccurate focal length estimate or tilt angle), we decided to use local repre-
sentations of the skyline (see [38] for an overview on shape features).
To describe the contour, we consider overlapping curvelets of width w(imagine
a sliding window, see Fig. 1). These curvelets are then sampled at nequally spaced
Image Based Geo-Localization in the Alps 7
(a)
(b)
(c)
(d)
(e)
(f)
-0.14 -0.01 0.10 0.17 0.15 0.04 -0.10 -0.21
001 011 101 111 110 100 001 000
001011101111110100001000
Fig. 3 Contour word computation: (a) raw contour, (b) smoothed contour with nsampled points, (c) sam-
pled points after normalization, (d) contourlet as numeric vector, (e) each dimension quantized to 3 bits,
(f) contour word as 24-bit integer.
points, yielding each an n-dimensional vector ˜y1,..., ˜yn(before sampling, we low-
pass filter the skyline to avoid aliasing). The final descriptor is obtained by subtracting
the mean and dividing by the feature width (see Fig. 3(a)–(d)):
yi=˜yi−¯y
wfor i=1,...,nwhere ¯y=1
n
n
∑
j=1
˜yj(4)
Mean subtraction makes the descriptor invariant w.r.t. vertical image location (and
therefore robust against camera tilt). Scaling ensures that the yi’s have roughly the
same magnitude, independently of the feature width w.
In a next step, each dimension of a contourlet is quantized (Fig. 3(e)–(f)). Since
the features are very low-dimensional compared to traditional patch-based feature
descriptors like SIFT [21], we choose not to use a vocabulary tree. Instead, we directly
quantize each dimension of the descriptor separately, which is both faster and more
memory-efficient compared to a traditional vocabulary tree. In addition the best bin
is guaranteed to be found. Each yifalls into one bin and the nassociated bin numbers
are concatenated into a single integer, which we refer to as contour word. For each
descriptor, the viewing direction αq, relative to the camera’s optical axis is computed
using the camera’s intrinsics parameters and is stored together with the visual word.
We have verified that an approximate focal length estimate is sufficient. In case of an
unknown focal length, it is possible to sample several tentative focal length values,
which we evaluate in Section 3.
2.2 Visual Database Creation
The digital elevation model we use for validation is available from the Swiss Federal
Office of Topography, and similar datasets exist also for the US and other countries.
There is one sample point per 2 square meters and the height quality varies from 0.5m
(flat regions) to 3m-8m (above 2000m elevation) average error1. This data is con-
verted to a triangulated surface model with level-of-detail support in a scene graph
representation2. At each position on a regular grid on the surface (every 0.001◦in
1http://www.swisstopo.admin.ch/internet/swisstopo/en/home
2http://openscenegraph.org
8 Olivier Saurer et al.
N-S direction and 0.0015◦in E-W direction, i.e. 111m and 115m respectively) and
from 1.80m above the ground3, we render a cube-map of the textureless DEM (face
resolution 1024×1024) and extract the visible skyline by checking for the rendered
sky color. Overall, we generate 3.5 million cubemaps. Similar to the query image,
we extract contourlets, but this time with absolute viewing direction. We organize
the contourlets in an index to allow for fast retrieval. In image search, inverted files
have been used very successfully for this task [31]. We extend this idea by also taking
into account the viewing direction, so that we can perform rough geometric verifica-
tion on-the-fly. For each word we maintain a list that stores for every occurrence the
panorama ID and the azimuth αdof the contourlet.
2.3 Recognition and Verification
2.3.1 Baseline
The baseline for comparison is an approach borrowed from patch based systems (e.g.
[25,28,8]) based on the (potentially weighted) L1-norm between normalized visual
word frequency vectors:
DE(˜
q,˜
d) = k˜
q−˜
dk1=∑
i
|˜qi−˜
di|or DEw(˜
q,˜
d) = ∑
i
wi|˜qi−˜
di|(5)
with ˜
q=q
kqk1
and ˜
d=d
kdk1
(6)
Where qiand diis the number of times visual word iappears in the query or database
image respectively, and ˜qi,˜
diare their normalized counterparts. wiis the weight of
visual word i(e.g. as obtained by the term frequency - inverse document frequency
(tf-idf) scheme). This gives an ideal score of 0 when both images contain the same
visual words at the same proportions, which means that the L1-norm favors images
that are equal to the query.
Nister et al. [25] suggested transforming the weighted L1-norm like this
DEw(˜
q,˜
d) = ∑
i
wi˜qi+∑
i
wi˜
di−2∑
i∈Q
wimin(˜qi,˜
di)(7)
in order to enable an efficient method for evaluating it by iterating only over the visual
words present in the query image and updating only the scores of database images
containing the given visual word.
2.3.2 “Contains”-Semantics
In our setting, we are comparing 10◦–70◦views to 360◦panoramas, which means that
we are facing a 5×–36×difference of magnitude. Therefore, it seems ill-advised to
3Synthetic experiments verified that taking the photo from ten or fifty meters above the ground does
not degrade recognition besides very special cases like standing very close to a small wall.
Image Based Geo-Localization in the Alps 9
implement an “equals”-semantics, but rather one should use a “contains”-semantics.
We modify the weighted L1-norm as follows:
DC(q,d) = ∑
i
wimax(qi−di,0)(8)
The difference is that we are using the raw contour word frequencies, qiand diwith-
out scaling and we replace the absolute value |·| by max(·,0). Therefore, one only
penalizes contour words that occur in the query image, but not in the database image
(or more often in the query image than in the database image). An ideal score of 0
is obtained by a database image that contains every contour word at least as often as
the query image, plus any number of other contour words. If the proposed score is
transformed as follows, it can be evaluated just as efficiently as the baseline:
DC(q,d) = ∑
i∈Q
wiqi−∑
i∈Q
wimin(qi,di)(9)
This subtle change makes a huge difference, see Fig. 6(a) and Table 1: (B) versus (C).
Note that this might also be applicable to other cases where a “contains”-semantics
is desirable.
2.3.3 Location and Direction
We further refine retrieval by taking geometric information into account already dur-
ing the voting stage. Earlier bag-of-words approaches accumulate evidence purely
based on the frequency of visual words. Voting usually returns a short-list of the top
ncandidates, which are reranked using geometric verification (typically using the
number of geometric inliers). For performance reasons, nhas to be chosen relatively
small (e.g. n=50). If the correct answer already fails to be in this short-list, then
no amount of reordering can bring it back. Instead, we check for geometric consis-
tency already at the voting stage, so that fewer good candidates get lost prematurely.
Not only does this increase the quality of the short-list, it also provides an estimated
viewing direction, which can be used as an initial guess for the full geometric verifica-
tion. Since this enables a significant speedup, we can afford to use a longer short-list,
which further reduces the risk of missing the correct answer.
If the same contour word appears in the database image at angle αd(relative to
north) and in the query image at angle αq(relative to the camera’s optical axis), the
camera’s azimuth can be calculated as α=αd−αq. Weighted votes are accumulated
using soft binning and the most promising viewing direction(s) are passed on to full
geometric verification. This way, panoramas containing the contour words in the right
order get many votes for a single direction, ensuring a high score. For panoramas con-
taining only the right mix of contour words, but in random order, the votes are divided
among many different directions, so that none of them gets a good score (see Fig. 4).
Note that this is different from merely dividing the panoramas into smaller sections
and voting for these sections: Our approach effectively requires that the order of con-
tour words in the panorama matches the order in the query image. As an additional
benefit, we do not need to build the inverted file for any specific field-of-view of the
query image.
10 Olivier Saurer et al.
(a) (b) (c) (d) (e)
w1
w3
w2
β1
β2
β3
w1
w3
w2
α1
α2
α3
w1
w3
w2
α1
α2
α3
w1
w3
w2
α
w1
w3
w2
α1
α2
α3
Optical Axis
N
N
N
N
Fig. 4 Voting for a direction is illustrated using a simple example: We have a query image (a) with contour
words wiand associated angles βirelative to the optical axis. We consider a panorama (b) with contour
words in the same relative orientation αias the query image. Since the contour words appear in the same
order, they all vote for the same viewing direction α(c). In contrast, we consider a second panorama (d)
with contour words in a different order. Even though the contour words occur in close proximity they each
vote for a different direction αi, so that none of the directions gets a high score (e).
2.3.4 Geometric Verification
After retrieval we geometrically verify the top 1000 candidates. The verification con-
sists in computing an optimal alignment of the two visible skylines using iterative
closest points (ICP). While we consider in the voting stage only one angle (azimuth),
ICP determines a full 3D rotation. First, we sample all possible values for azimuth
and keep the two other angles at zero. The most promising one is used as initializa-
tion for ICP. In the variants that already vote for a direction, we try only a few values
around the highest ranked ones. The average alignment error is used as a score for
re-ranking the candidates.
3 Evaluation
In this section we evaluate the proposed algorithm on two real datasets consisting of
a total of 1151 images. We further give a detailed evaluation of the algorithm under
varying tilt and roll angles, and show that in cases where the focal length parameter
is unknown it can effectively be sampled.
Query Set.
In order to evaluate the approaches we assembled two datasets, which we refer to as
CH1 and CH2. The CH1 dataset consists of 203 photographs obtained from different
sources such as online photo collections and on site image capturing. The CH2 dataset
consists of 948 images which were solely captured on site. For all of the photographs,
we verified the GPS tag or location estimate by comparing the skyline to the surface
model. For the majority of the images the information was consistent. For a few
of them the position did not match the digital elevation model’s view. This can be
explained by a wrong cell phone GPS tag, due to bad/no GPS reception at the time
the image was captured. For those cases, we use dense geometric verification (on
each 111m×115m grid position up to a 10km radius around the tagged position) to
generate hypotheses for the correct GPS tag. We verify this by visual inspection and
removed images in case of disagreement. The complete set of query images used
Image Based Geo-Localization in the Alps 11
Fig. 5 Oblique view of Switzerland, spanning a total 40 000km2. Spheres indicate the query images’ of the
CH1 (red) and CH2 (blue) dataset at ground truth coordinates (size reflects 1km tolerance radius). Source
of DEM: Bundesamt für Landestopografie swisstopo (Art. 30 GeoIV): 5704 000 000
is available at the project website4. The distribution of the CH1 and CH2 dataset
is drawn on to the DEM in Fig. 5. For all of the query images FoV information is
available (e.g. from EXIF tag). However, we have verified experimentally that also in
case of fully unknown focal length the system can be applied by sampling over this
parameter, see Fig. 10 as example and subsection 3.
Query Image Segmentation.
We used the CH1 query images which were already segmented in [2] as training set
and apply our segmentation pipeline to the CH2 dataset. Out of the 948 image 60% of
the images were segmented fully automatically, while 30% required little user inter-
action, mainly to correct for occluders such as trees or buildings. 10% of the images
required a more elaborate user interaction, to correct for snow fields, (often confused
as sky), clouds hiding small parts of the mountain or for reflections appearing when
taking pictures from inside a car, cable-car or train. Our new segmentation pipeline
improved by 18%, compared to the previous method proposed in [2].
Parameter Selection.
The features need to be clearly smaller than the images’ field-of-view, but wide
enough to capture the geometry rather than just discretization noise. We consider de-
scriptors of width w=10◦and w=2.5◦. The number of sample points nshould not be
so small that it is uninformative (e.g. n=3 would only distinguish concave/convex),
but not much bigger than that otherwise it risks being overly specific, so we choose
n=8. The curve is smoothed by a Gaussian with σ=w
2n, i.e. half the distance be-
tween consecutive sample points. Descriptors are extracted every σdegrees.
Each dimension of the descriptor is quantized into kbins of width 0.375, the first
and last bin extending to infinity. We chose kas a power of 2 that results in roughly
1 million contour words, i.e. k=8. This maps each yito 3 bits, producing contour
4http://cvg.ethz.ch/research/mountain-localization
12 Olivier Saurer et al.
words that are 24 bit integers. Out of the 224 potential contour words, only 300k–
500k (depending on w) remain after discarding words that occur too often (more than
a million) or not at all.
Recognition Performance.
The recognition pipeline using different voting schemes and varying descriptor sizes
is evaluated on both datasets, see Table 1. All of the tested recognition pipelines return
a ranked list of candidates. We evaluate them as follows: For every n=1,...,100,
we count the fraction of query images that have at least one correct answer among
the top ncandidates. We consider an answer correct if it is within 1km of the ground
truth position (see Fig. 6).
Number of Candidates
0 20 40 60 80 100
Fraction of Correctly Localized Images
0
0.2
0.4
0.6
0.8
1
(E) Location and direction
(C) "contains"-semantics
(B) "equals"-semantics
(A) Random guessing
Number of Candidates
0 20 40 60 80 100
Fraction of Correctly Localized Images
0
0.2
0.4
0.6
0.8
1
(D) 2
°
(E) 3
°
(F) 5
°
(G) 10
°
(H) 20
°
Number of Candidates
0 20 40 60 80 100
Fraction of Correctly Localized Images
0
0.2
0.4
0.6
0.8
1
(E) Location and direction
(C) "contains"-semantics
(B) "equals"-semantics
(A) Random guessing
Number of Candidates
0 20 40 60 80 100
Fraction of Correctly Localized Images
0
0.2
0.4
0.6
0.8
1
(D) 2
°
(E) 3
°
(F) 5
°
(G) 10
°
(H) 20
°
(a) (b)
Fig. 6 Retrieval performance for different: (a) voting schemes, (b) bin sizes in direction voting. Evaluated
on the CH1 (top) and CH2 (bottom) dataset.
In Fig. 6(a), we compare different voting schemes: (B) voting for location only,
using the traditional approach with normalized visual word vectors and L1-norm
(“equals”-semantics); (C) voting for location only, with our proposed metric (“contains”-
semantics); (E) voting for location and direction simultaneously (i.e. taking order
into account). All variants use 10◦descriptors. For comparison, we also show (A)
the probability of hitting a correct panorama by random guessing (the probability
of a correct guess is extremely small, which shows that the tolerance of 1km is not
overly generous). Our proposed “contains”-semantics alone already outperforms the
baseline (“equals”-semantics) by far, but voting for a direction is even better.
Image Based Geo-Localization in the Alps 13
Voting scheme Descriptor width Dir. bin size Geo. ver. CH1 (top 1 corr.) CH2 (top 1 corr.)
(A) random N/A N/A no 0.008% 0.008%
(B) “equals” 10◦N/A no 9% 1%
(C) “contains” 10◦N/A no 31% 21%
(D) loc.&dir. 10◦2◦no 45% 30%
(E) loc.&dir. 10◦3◦no 43% 31%
(F) loc.&dir. 10◦5◦no 46% 31%
(G) loc.&dir. 10◦10◦no 42% 30%
(H) loc.&dir. 10◦20◦no 38% 28%
(I) loc.&dir. 2.5◦3◦no 28% 14%
(J) loc.&dir. 10◦&2.5◦3◦no 62% 44%
(K) loc.&dir. 10◦&2.5◦3◦yes 88% 76%
Table 1 Overview of tested recognition pipelines.
In Fig. 6(b), we analyse how different bin sizes for direction voting affects re-
sults. (D)–(H) correspond to bin sizes of 2◦
,3◦
,5◦
,10◦
,20◦respectively. While there
are small differences, none of the settings outperforms all others consistently: Our
method is quite insensitive over a large range of this parameter.
In Fig. 7(a), we study the impact of different descriptor sizes: (E) only 10◦de-
scriptors; (I) only 2.5◦descriptors; (J) both 10◦and 2.5◦descriptors combined. All
variants vote for location and direction simultaneously. While 10◦descriptors outper-
forms 2.5◦descriptors, the combination of both is better than either descriptor size
alone. This demonstrates that different scales capture different information, which
complement each other.
In Fig. 7(b), we show the effect of geometric verification by aligning the full
countours using ICP: (J) 10◦and 2.5◦descriptors voting for location and direction,
without verification; (K) same as (J) but with geometric verification. We see that ICP
Number of Candidates
0 20 40 60 80 100
Fraction of Correctly Localized Images
0
0.2
0.4
0.6
0.8
1
(J) 10
°
and 2.5
°
(E) 10
°
only
(I) 2.5
°
only
Number of Candidates
0 20 40 60 800 1000
Fraction of Correctly Localized Images
0
0.2
0.4
0.6
0.8
1
(K) Geometric verification
(J) Without GV
Distance to Ground Truth Position (km)
0 1 2 3 4 5
Fraction of Correctly Localized Images
0
0.2
0.4
0.6
0.8
1
Distribution of error
Cumulative distribution
Number of Candidates
0 20 40 60 80 100
Fraction of Correctly Localized Images
0
0.2
0.4
0.6
0.8
1
(J) 10
°
and 2.5
°
(E) 10
°
only
(I) 2.5
°
only
Number of Candidates
0 20 40 60 800 1000
Fraction of Correctly Localized Images
0
0.2
0.4
0.6
0.8
1
(K) Geometric verification
(J) Without GV
Distance to Ground Truth Position (km)
0 1 2 3 4 5
Fraction of Correctly Localized Images
0
0.2
0.4
0.6
0.8
1
Distribution of error
Cumulative distribution
(a) (b) (c)
Fig. 7 Retrieval performance for CH1 (top) and CH2 (bottom) dataset: (a) Different descriptor sizes. (b)
Retrieval performance before and after geometric verification. (c) Fraction of queries having at most a
given distance to the ground truth position. Not shown: 21 images (9.9%) from the CH1 dataset with an
error between 7 and 217km and 177 images (18.6%) from the CH2 dataset with an error between 13 and
245km.
14 Olivier Saurer et al.
based reranking is quite effective at moving the best candidate(s) to the beginning
of the short list: On the CH1 dataset the top ranked candidate is within a radius of
1km with a probability of 88%. On the CH2 dataset we achieve a recognition rate of
76% for a maximum radius of 1km. See Fig. 7(c) for other radii. In computer assisted
search scenarios, an operator would choose an image from a small list which would
further increase the percentage of correctly recovered pictures. Besides that, from
geometric verification we not only obtain an estimate for the viewing direction but
the full camera orientation which can be used for augmented reality. Fig. 8 and 9
show images of successful and unsuccessful localization.
Field-of-View.
In Fig. 10 we illustrate the effect of inaccurate or unknown field-of-view (FoV). For
one query image, we run the localization pipeline (K) assuming that the FoV is 11◦
and record the results. Then we run it again assuming that the FoV is 12◦etc., up
to 70◦. Fig. 10 shows how the alignment error and estimated position depend on the
assumed FoV.
In principle, it is possible to compensate a wrong FoV by moving forward or
backward. This holds only approximately if the scene is not perfectly planar. In ad-
dition, the effect has hard limits because moving too far will cause objects to move
in or out of view, changing the visible skyline. Between these limits, changing the
FoV causes both the alignment error and the position to change smoothly. Outside of
this stable range, the error is higher, fluctuates more and the position jumps around
wildly.
This has two consequences: First, if the FoV obtained from the image’s metadata
is inaccurate it is usually not a disaster, the retrieved position will simply be slightly
inaccurate as well, but not completely wrong. Second, if the FoV is completely un-
known, one can get a rough estimate by choosing the minimum error and/or looking
for a range where the retrieved position is most stable.
The field-of-view (FoV) extracted from the EXIF data may not always be 100%
accurate. This experiment studies the effects of a slight inaccuracy. We modify the
FoV obtained from the EXIF by ±5% and plot it against the recognition rate obtained
over the entire query set CH1. We observe in Fig. 11(a) that even if the values are off
by ±5%, we still obtain a recognition rate of 70 −80%.
Tilt Angle.
Our algorithm assumes that landscape images usually are not subject to extreme tilt
angles. In the final experiment evaluated in Fig. 11(b), we virtually rotate the ex-
tracted skyline of the query images by various angles in order to simulate camera tilt
and observe how recognition performance is affected. As shown in Fig. 11(b) with
30◦tilt we still obtain a recognition rate of 60% on the CH1 dataset. This is a large
tilt angle, considering that the skyline is usually straight in front of the camera and
not above or below it.
Image Based Geo-Localization in the Alps 15
Roll Angle.
Our algorithm makes a zero roll assumption, meaning that the camera is held upright.
To evaluate the robustness of the algorithm we virtually perturb the roll angle by ro-
tating the extracted skyline of the query image by various angles. Fig. 11(c) shows
the achieved recognition rate. For 5◦roll angle the recognition rate drops by 26%.
This drop does not come as a surprise since the binning of the skyline makes a strong
assumption on a upright image. In general this assumption can be relaxed by extend-
ing the database with differently rotated skylines, or by using IMU data (often present
in today’s mobile phones) to correct for the roll angle in the query image. In general
we found that landscape images captured with a hand held camera are subject to very
little roll rotation, which is also confirmed by both datasets.
Runtime.
We implemented the algorithm partly in C/C++ and partly in Matlab. The segmen-
tation runs at interactive frame rate and gives direct visual feedback to the operator,
given the unary potential of our segmentation framework. Given the skyline it takes
10 seconds to find the camera’s position and rotation in an area of 40000km2per
image. Exhaustively computing an optimal alignment between the query image and
each of the 3.5M panoramas would take on the order of several days. For comparison,
the authors of [3] use a GPU implementation and report 2 minutes computation time
to determine the rotation only, assuming the camera position is already known.
4 Conclusion and Future Work
We have presented a system for large scale location recognition based on digital el-
evation models. This is very valuable for geo-localization of pictures when no GPS
information is available (for virtually all video or DSLR cameras, archive pictures,
in intelligence and military scenarios). We extract the sky and represent the visible
skyline by a set of contour words, where each contour word is represented together
with its offset angle from the optical axis. This way, we can do a bag-of-words like
approach with integrated geometric verification, i.e. we are looking for the panorama
(portion) that has a similar frequency of contour words with a consistent direction.
We show that our representation is very discriminative and the full system allows for
excellent recognition rates on the two challenging dataset. On the CH1 dataset we
achieve a recognition rate of 88% and 76% on the CH2 dataset. Both datasets include
different seasons, landscapes and altitudes. We believe that this is a step towards the
ultimate goal of being able to geo-localize images taken anywhere on the planet, but
for this also other additional cues of natural environments have to be combined with
the given approach. This will be the subject of future research.
Acknowledgements This work has been supported through SNF grant 127224 by the Swiss National Sci-
ence Foundation. We also thank Simon Wenner for his help to render the DEMs and Hiroto Nagayoshi for
providing the CH2 dataset. We also thank the anonymous reviewers for useful discussions and constructive
feedback.
16 Olivier Saurer et al.
References
1. G. Baatz, K. Köser, D. Chen, R. Grzeszczuk, and M. Pollefeys. Leveraging 3d city models for rotation
invariant place-of-interest recognition. International Journal of Computer Vision (IJCV), Special
Issue on Mobile Vision, 96, 2012.
2. G. Baatz, O. Saurer, K. Köser, and M. Pollefeys. Large scale visual geo-localization of images in
mountainous terrain. In Proceedings of European Conference on Computer Vision (ECCV), pages
517–530, 2012.
3. L. Baboud, M. Cadík, E. Eisemann, and H.-P. Seidel. Automatic photo-to-terrain alignment for the
annotation of mountain pictures. In Proceedings of Computer Vision and Pattern Recognition (CVPR),
pages 41–48, 2011.
4. M. Bansal and K. Daniilidis. Geometric urban geo-localization. In Proceedings of Computer Vision
and Pattern Recognition (CVPR), pages 3978–3985, 2014.
5. J.-C. Bazin, I. Kweon, C. Demonceaux, and P. Vasseur. Dynamic programming and skyline extrac-
tion in catadioptric infrared images. In Proceedings of International Conference on Robotics and
Automation (ICRA), pages 409–416, 2009.
6. A. Blake, C. Rother, M. Brown, P. Perez, and P. Torr. Interactive image segmentation using an adaptive
gmmrf model. In Proceedings of European Conference on Computer Vision (ECCV), pages 428–441,
2004.
7. M. Brown and D. G. Lowe. Automatic panoramic image stitching using invariant features. Interna-
tional Journal of Computer Vision (IJCV), 74:59–73, August 2007.
8. D. Chen, G. Baatz, Köser, S. Tsai, R. Vedantham, T. Pylvanainen, K. Roimela, X. Chen, J. Bach,
M. Pollefeys, B. Girod, and R. Grzeszczuk. City-scale landmark identification on mobile devices. In
Proceedings of Computer Vision and Pattern Recognition (CVPR), 2011.
9. D. Comaniciu, P. Meer, and S. Member. Mean shift: A robust approach toward feature space analysis.
IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 24:603–619, 2002.
10. F. Cozman. Decision Making Based on Convex Sets of Probability Distributions: Quasi-Bayesian
Networks and Outdoor Visual Position Estimation. PhD thesis, Robotics Institute, Carnegie Mellon
University, Pittsburgh, PA, December 1997.
11. F. Cozman and E. Krotkov. Position estimation from outdoor visual landmarks for teleoperation of
lunar rovers. In WACV ’96, pages 156 –161, 1996.
12. J. Friedman, T. Hastie, and R. Tibshirani. Additive Logistic Regression: a Statistical View of Boosting.
The Annals of Statistics, 2000.
13. J. Hays and A. A. Efros. im2gps: estimating geographic information from a single image. In Pro-
ceedings of Computer Vision and Pattern Recognition (CVPR), 2008.
14. S. u. Hussain and B. Triggs. Visual recognition using local quantized patterns. In Proceedings of
European Conference on Computer Vision (ECCV), 2012.
15. V. Kolmogorov and Y. Boykov. What metrics can be approximated by geo-cuts, or global optimization
of length/area and flux. In Proceedings of International Conference on Computer Vision (ICCV),
pages 564–571, Washington, DC, USA, 2005.
16. L. Ladicky, C. Russell, P. Kohli, and P. Torr. Associative hierarchical random fields. Pattern Analysis
and Machine Intelligence (PAMI), 36(6):1056–1077, June 2014.
17. L. Ladicky, B. Zeisl, and M. Pollefeys. Discriminatively trained dense surface normal estimation. In
Proceedings of European Conference on Computer Vision (ECCV), 2014.
18. J.-F. Lalonde, S. G. Narasimhan, and A. A. Efros. What do the sun and the sky tell us about the
camera? International Journal on Computer Vision, 88(1):24–51, May 2010.
19. Y. Li, N. Snavely, and D. P. Huttenlocher. Location recognition using prioritized feature matching. In
Proceedings of European Conference on Computer Vision (ECCV), pages 791–804, 2010.
20. W.-N. Lie, T. C.-I. Lin, T.-C. Lin, and K.-S. Hung. A robust dynamic programming algorithm to
extract skyline in images for navigation. Pattern Recognition Letters, 26(2):221 – 230, 2005.
21. D. G. Lowe. Distinctive Image Features from Scale-Invariant Keypoints. International Journal of
Computer Vision (IJCV), 60(2):91–110, 2004.
22. J. Malik, S. Belongie, T. Leung, and J. Shi. Contour and texture analysis for image segmentation.
International Journal of Computer Vision, 43(1):7–27, June 2001.
23. S. Manay, D. Cremers, B.-W. Hong, A. Yezzi, and S. Soatto. Integral invariants for shape matching.
Pattern Analysis and Machine Intelligence (PAMI), 2006.
24. P. C. Naval, M. Mukunoki, M. Minoh, and K. Ikeda. Estimating camera position and orientation from
geographical map and mountain image. In 38th Pattern Sensing Group Research Meeting, Soc. of
Instrument and Control Engineers, pages 9–16, 1997.
25. D. Nistér and H. Stewénius. Scalable recognition with a vocabulary tree. In Proceedings of Computer
Vision and Pattern Recognition (CVPR), pages 2161–2168, 2006.
Image Based Geo-Localization in the Alps 17
26. S. Ramalingam, S. Bouaziz, and P. Sturm. Pose estimation using both points and lines for geo-
localization. In Proceedings of International Conference on Robotics and Automation (ICRA), pages
4716–4723, 2011.
27. S. Ramalingam, S. Bouaziz, P. Sturm, and M. Brand. Skyline2gps: Localization in urban canyons
using omni-skylines. In IROS 2010, pages 3816 –3823, oct. 2010.
28. G. Schindler, M. Brown, and R. Szeliski. City-scale location recognition. In Proceedings of Computer
Vision and Pattern Recognition (CVPR), pages 1 –7, june 2007.
29. E. Shechtman and M. Irani. Matching local self-similarities across images and videos. In Proceedings
of Conference on Computer Vision and Pattern Recognition (CVPR), 2007.
30. J. Shotton, J. Winn, C. Rother, and A. Criminisi. Textonboost: Joint appearance, shape and context
modeling for multi-class object recognition and segmentation. In Proceedings of European Confer-
ence on Computer Vision (ECCV), pages 1–15, 2006.
31. J. Sivic and A. Zisserman. Video Google: A text retrieval approach to object matching in videos. In
Proceedings of International Conference on Computer Vision (ICCV), pages 1470–1477, oct 2003.
32. F. Stein and G. Medioni. Map-based localization using the panoramic horizon. Transaction on
Robotics and Automation, 11(6):892 –896, dec 1995.
33. R. Talluri and J. Aggarwal. Position estimation for an autonomous mobile robot in an outdoor envi-
ronment. Transaction on Robotics and Automation, 8(5):573 –584, oct 1992.
34. A. Taneja, L. Ballan, and M. Pollefeys. Registration of spherical panoramic images with cadastral 3d
models. In 3D Imaging, Modeling, Processing, Visualization and Transmission (3DIMPVT), pages
479–486, 2012.
35. W. B. Thompson, T. C. Henderson, T. L. Colvin, L. B. Dick, and C. M. Valiquette. Vision-based
localization. In Image Understanding Workshop, pages 491–498, 1993.
36. A. Vasilevskiy and K. Siddiqi. Flux maximizing geometric flows. Transactions on Pattern Analysis
and Machine Intelligence (PAMI), pages 1565–1578, 2002.
37. J. Woo, K. Son, T. Li, G. S. Kim, and I.-S. Kweon. Vision-based uav navigation in mountain area. In
MVA, pages 236–239, 2007.
38. M. Yang, K. Kpalma, and J. Ronsin. A Survey of Shape Feature Extraction Techniques. In P.-Y. Yin,
editor, Pattern Recognition, pages 43–90. IN-TECH, Nov 2008.
18 Olivier Saurer et al.
Fig. 8 Sample Results: First and fourth column are input images. Second and fifth column show the
segmentations and third and sixth column show the query images augmented with the skyline, retrieved
from the database. The images in the last five rows were segmented with help of user interaction.
Image Based Geo-Localization in the Alps 19
Fig. 9 Some incorrectly localized images. This usually happens to images with a relatively smooth skyline
and only few distinctive features. The pipeline finds a contour that fits somewhat well, even if the location
is completely off.
10 20 27 33 42 50 60 70
1
2
3
4
5
6
7
8
FoV (degrees)
Alignment Error (pixels)
Northing (km)
Easting (km)
11
27
33
42
70
−5 0 5
−5
0
5
0
5
10
15
Northing (km)
Easting (km)
11
27
33
42
70
−5 0 5
−5
0
5
(a) (b) (c) (d)
Fig. 10 (a) Query image. (b) Alignment error of the best position for a given FoV. Dashed lines indi-
cate the limits of the stable region and the FoV from the image’s EXIF tag. (c) Alignment error of the
best FoV for a given position. For an animated version, see http://cvg.ethz.ch/research/
mountain-localization. (d) Shaded terrain model. The overlaid curve in (c) and (d) starts from the
best location assuming 11◦FoV and continues to the best location assuming 12◦, 13◦, etc. Numbers next
to the markers indicate corresponding FoV.
Change in Field-of-View (percent)
-5 0 5
Fraction of Correctly Localized Images
0
0.2
0.4
0.6
0.8
1
Additional Tilt (degrees)
0 10 20 30
Fraction of Correctly Localized Images
0
0.2
0.4
0.6
0.8
1
Additional Roll (degrees)
0 1 2 3 4 5
Fraction of Correctly Localized Images
0
0.2
0.4
0.6
0.8
1
Change in Field-of-View (percent)
-5 0 5
Fraction of Correctly Localized Images
0
0.2
0.4
0.6
0.8
1
Additional Tilt (degrees)
0 10 20 30
Fraction of Correctly Localized Images
0
0.2
0.4
0.6
0.8
1
Additional Roll (degrees)
0 1 2 3 4 5
Fraction of Correctly Localized Images
0
0.2
0.4
0.6
0.8
1
(a) (b) (c)
Fig. 11 Robustness evaluation under: (a) varying FoV, (b) varying tilt angle, (c) varying roll angle. Top
row CH1 and bottom row CH2 dataset.
