Do geographic features impact pictures location shared on the Web? Modeling photographic suitability in the Swiss Alps. Produit Timothée1 , Tuia Devis1 , de Morsier Frank2 , Golay François1 1 LaSIG laboratory 2 Signal Processing Laboratory (LTS5) Ecole Polytechnique Fédérale de Lausanne 1015 Lausanne Switzerland ABSTRACT Keywords Nowadays, millions of landscape images are uploaded on Geographic One-Class data (GOCD), image geotags, land- photosharing platforms such as Flickr or Panoramio. More scape images, density mapping and more of these images are also accurately geotagged via GPS devices mounted on personal cameras. Each image re- sults from a twofold spatial choice: the choice of the location and the choice of the picture subject. 1. INTRODUCTION In this study, our focus is on landscape images in large With the advent of ”Web 2.0”, the number of collabora- touristic areas. Firstly, our goal is to learn which geographic tive databases has dramatically increased. Very often, the features play a role in the choice of the location of shared databases objects are geotagged, meaning that an attribute images. For our analysis, we extract a series of geographic related to its geographic location is available. features from a Digital Elevation Model (DEM) and a To- In this paper, we focus on collective pictures databases. pographic Landscape Model (TLM) and model the photo- Pictures uploaded on the Web via a photosharing platform graphic suitability as a density estimation problem in the (Flickr, Panoramio, Instagram...) have a time stamp, some space of the geographic features. Secondly, each combina- textual tags describing its content and sometimes a world tion of geographic features of a region is associated with a coordinate representing the geographic position. We are probability to be a location suitable for a photography. The specially interested to the geotag which can be attributed resulting map is useful to promote tourism, to evaluate the in several ways. Most of the images are located with a click landscape attractiveness or with a more technical objective on a map. The accuracy of the localization is related to the as a prior in close-range photogrammetry. zoom level used and to the ability to recognize an area in This study shows that databases of geotagged pictures can an aerial view. On the contrary, more and more cameras be used to understand tourists behaviour also in rural ar- have a GPS device able to track very accurately the latitude eas, even if most of current researches are adressed to cities. and longitude coordinates. As stated in [7], the location The application to a touristic region in the Swiss Alps shows of a picture is the result of a spatial choices: the choice of that the proposed method suits well this Geographic One- the location (the place where the photograph stands) and Class Data problem and is more accurate than both stan- the choice of the subject (the object at which the camera dard KPCA and One-Class SVM to model the suitability points). This study focuses on the first choice: it benefits for touristic photography at locations unseen during train- from the GPS accuracy to learn which geographic features ing. As expected, picture locations are mostly correlated describe the locations chosen. with geographic features extracted from the digital elevation Such a study aims at providing a map of the suitability to model, as well as with those related to accessibility (distance be picture location. This map can be useful in several ways. to roads, paths). However, the force of this study is the com- First, it could be used to promote tourism in areas sharing bination of the geographic features via a kernel method to similar geographic feature with famous regions. Second, it model more accurately suitable picture locations. can be used to assess landscape attractiveness, a measure needed in environmental planning [13]. Finally, shared pho- tographies in landscape areas could be a valuable source to Categories and Subject Descriptors extract information about natural phenomena (displacement H.2.8 [Database management]: Database Applications of glaciers [18], landscape change [3], local meteorology...): Spatial databases and GIS to meet this goal they require to be located and oriented ac- curately. Recent research shows that landscape images pose estimation (computation of the location and orientation of Copyright ° c by the paper’s authors. Copying permitted only for private a camera in the computer vision vocabulary) requires priors and academic purposes. which can be provided by GIS data: horizon and 3D models In: S. Vrochidis, K. Karatzas, A. Karpinnen, A. Joly (eds.): Proceedings of the International Workshop on Environmental Multimedia Retrieval (EMR [1, 2, 4], aerial views [17], sun position [11]. The proposed 2014), Glasgow, UK, April 1, 2014, published at http://ceur-ws.org map can thus be used to extract the most probable picture 22 locations in a neighborhood. such outliers are present, some of the labelled data belong However, the collaborative database are more widely used to the negative class. to analyse people behaviour and general trends in the tourists We will therefore focus on non-parametric methods for movements in urban areas [23, 16, 7]. Picture locations distribution support estimation. Such methods do not re- drawn on a map are difficult to read and require a more ap- quire knowledge about the distribution of the data and fit propriate geovisualization. In [16, 7], spatial density maps well with our problem. Besides support vector methods, (also called heat maps) are used to extract easily the most another straightforward method is the Kernel Density Esti- attractive areas (see for instance the site sightsmap.com). mation (KDE, also known as Parzen window) [15, 19]: KDE The textual geotags associated to a world coordinate also estimates the density of data by applying a local smoothing give many opportunities. For instance, authors in [8] use filter [26]. the tags to draw the geographic boundaries of fuzzy regions. In this paper, we propose a KDE-based strategy to es- They compare how Kernel Density Estimation (KDE) and timate the probability distribution function (PDF) of the Support Vector Machines (SVM) are accurate in the extrac- data. We estimate the support of the data both in the orig- tion of areas such as the Alps. Heat and density maps are inal space of geographic features and in the feature space well suited for large scale mapping. However, once zooming spanned by KPCA. The PDFs of the labeled vs the unla- in, the contours of the high density regions become inaccu- beled data are used to define a Bayesian criterion, which rate, mainly because of the inaccuracy of the clicked geotags measures the probability for a map location to be an image and the use of a smoothing radius. For instance, locations location (given its geographic features). To fit the free pa- such cliffs can be considered highly attractive for photogra- rameters of KDE and KPCA we define a performance mea- phers just because they are within the influence radius of sure to ensure that most of the locations with photographies popular places, while, in reality, they are inaccessible. are classified as positives and most of the unlabeled loca- In order to interpolate probability values in each location tions are classified as negative. We compare the proposed of the map and not only in areas where picture are found, approaches with OCSVM and KPCA on a real dataset of we propose to compute density in a space generated from touristic pictures taken in the Swiss Alps. geographic features rather than the space of latitude and longitude. To this goal, we require precise image locations 2. METHODOLOGY which are provided by the GPS embedded in recent cameras and appropriate geographic features. Such geographic fea- 2.1 Geographic features extraction tures are extracted with a GI Software either from a Digital Elevation Model (DEM) or from a Topographic Landscape The geographic features are extracted using a GIS soft- Model (TLM: roads, lakes, forests). ware for an ensemble of N cells on a grid with 100m reso- By modeling the density of pictures in the space of ge- lution. The considered features are summarized in table 1 ographic features, we estimate the probability of being a and are computed for each cell zj of the grid forming the picture location over all the territory considered. This set- unlabeled set zu = {zj }N i=1 and for the n image locations ting is equivalent to a One-Class problem (OC) where there zi , forming the labeled set zl = {zi }n i=1 . Each zi and zj are is a set of positive data but no negative data: the locations thus represented by a d-dimensional vector formed by the d of landscape photographies compose the positive set, the set geographic features. of “attractive places”. However, for the map locations with The first set of features is extracted from the DEM (Alti- no pictures, we don’t know if the absence of photographies tude, slope, curvature and visible sky). Then, for the TLM is due to the inappropriateness of the location (a “bad” or based features, the distances from the cell to the nearest negative place) or simply to the lack of pictures in an “at- forest, lake, road and cable car are computed. All the geo- tractive” location. This type of problems is also common in graphic features are mean centered and scaled unit variance. geographic classification problems, such as change detection 2.2 GOCD with Kernel density Estimation from satellite images [14, 5]. Since in our case we consider a OC classification problem applied to geographical informa- The KDE function in equation (1) is used to evaluate the tion sciences, Guo [9] called these kinds of data Geographic density of a data set from the observations of the positive One-Class Data (GOCD). class zl . Once the density has been estimated, we can eval- During the last years, Kernel methods have been widely used uate the density for an unknown location zj . for OC problems. In this study, we consider the following 1 X ³ zj − zi ´ n OC models: the One-class SVM (OCSVM) [21], the Support fˆ(zj , zl ) = K (1) nh i=1 h Vector Data Description (SVDD) [24] and the Kernel PCA (KPCA) [10]. All of them use the kernel trick to project where K(x) is a local smoothing operator, or kernel func- the original data on a hypersphere in Rd . In the high di- tion. Among the different kernel functions, we used the mensional space, the data are more likely separated by a Gaussian function: linear model [20]. Guo [9] compares OCSVM, Maximum En- 1 tropy (MAXTENT) [12] and Positive and Unlabelled Learn- K(x) = (2π)−d/2 exp(− x2 ) (2) ing (PUL) [6] for GOCD problems. The two last methods 2 provided the most accurate results. However, in our case where h is the bandwidth of the kernel function. Scott’s study, the hypothesis which states that “the probability of rule is used to compute the bandwidth [22]: a negative data being labelled is null ” is not valid, thus ex- −1 cluding the use of PUL models. Indeed, our set of image h = n d+4 σzl (3) locations contains some outliers, for instance pictures taken where σzl is the covariance of the positive dataset. This from a cable car or pictures not related to landscapes. Since rule uses the dimension d of the dataset and the number 23 of positive data n to estimate a reasonable h. The choice similar, nevertheless both should differ from the random set of the appropriate kernel function has less influence on the density. Indeed, if the density of the random set and the one results than the choice of the proper bandwidth. Indeed, if issued from the labeled set are similar, the ratio for equation the bandwidth is too small, the density is over-fitted to the 5 will tend to one. This means that the geographic features positive data set and its generalization power is weak. On where badly chosen and are not able to distinguish properly the contrary, if the bandwidth is too large, the density will the random data from the labelled ones. In [9] the positive be oversmoothed and its small peaks will disappear. and unlabeled score presented in equation 6 is maximized: We consider the following scheme to describe our GOCD problem. Let Y ∈ 0, 1 be the event (or class) “is a picture r2 location”: Y = 1 for a cell being on a picture location and Fpu = (6) rpos Y = 0 otherwise. The probability of a cell being a “picture location”, given its geographic features zj , is Where r is the recall (the proportion of correctly predicted data in the testing set) and rpos is the ratio of positive pre- p(zj |Y = 1)P (Y = 1) dicted locations in the random set. P (Y = 1|zj ) = (4) p(zj ) In this study, we propose another criterion, more adapted P (Y = 1|zj ) is the value we want to compute for each cell of for the density estimation process: we want to ensure that the map. The data density p(zj ) is estimated with a KDE the bandwidth h fits well for both the training and testing sets and thereby their density should be similar. We want on a random set of unlabeled cells: fˆ(zj , zu ), the conditional to maximize C in equation 7: probability p(zj |Y = 1) is estimated from a KDE on the labeled data only: fˆ(zj , zl ) and P (Y = 1) is a unknown constant c. We are observing: s2R s2R C= = (7) sT − s2t 2 (sT − st )(sT + st ) P (Y = 1|zj ) p(zj |Y = 1) Where sR is 1 − rpos ; sT = (1 − rT ) with rT being the = (5) recall for the training set and st = (1 − rt ), with rt being the c p(zj ) recall for the testing set. Thus, C is very similar to Fpu but A threshold can be set on the cell probability from equa- the component (sT − st ) ensures that the PDF estimation tion 5 under which it is assumed to be drawn from the of the training and testing set are similar. generic distribution p(zj ), while above it is assumed drawn from the distribution of the labeled data p(zj , Y = 1). In 2.3 KDE and KDE(KPCA) practice, we set the threshold to one. In order to take into account the possible correlation be- To chose the best set of parameters for the method, a tween the geographic features and the non-linearity in the performance measure need to be defined. To this end, our data distribution, we propose to estimate the density of the labelled set is divided in three subsets, the first one to train data either in the original space or in a feature space spanned the KDE and the second to select the parameters. We want by the mapping φ(.) induced by the KPCA. Hoffmann [10] the estimated densities of the training and testing sets to be states that the density function is proportional to the spher- ical potential in the feature space. The spherical potential measures the distance between φ(z): the projection of the Table 1: Geographic features descriptions and ab- point z in the feature space, and the center of the data φ0 (z). breviations However, the spherical potential can’t be used to estimate the density of the unlabeled data, because if the kernel pro- Abbr. Description Source jection works well to separate the labeled set zl from the Z Altitude DEM, unlabeled set zu , their gravity centers φ0 (zl ) and φ0 (zu ) do Swisstopo not correspond. Consequently, we run the KDE on the non- Slope Slope in percent DEM, linear features extracted from KPCA. Swisstopo 2.4 Comparing approaches: One-Class SVM Curv Curvature DEM, and KPCA Swisstopo The kernel functions project the data on a hypersphere Sky Visible sky ratio, unobstructed DEM, which has the same dimension number than the number of hemisphere as a percentage. Swisstopo training data. OCSVM searches for a hyperplan with two DRoad Distance to the nearest road. TLM, constraints. First, the intersection between the plan and the Swisstopo sphere enclose a ratio of the points equal to 1 − µ, where µ BRoad Presence or absence of roads TLM, is the ratio of the outliers. Second, the plan has to be as (paths) in the cell. Swisstopo far as possible from the origin. SVDD [24] searches for the DFor Distance to the nearest forest TLM, minimal sphere which encloses the points. For “RBF” func- boundary, negative if the location Swisstopo tions, OCSVM and SVDD can produce similar results [10]. is within a forest. KPCA applies a PCA on the projection of the data on the sphere. The reconstruction error in the feature space is used DLake Distance to the nearest lake. TLM, to separate positive and negative data. It appears that this Swisstopo boundary encloses more tightly the data, resulting in best DLift Distance to the nearest cable car - TLM, results than OCSVM and SVDD. chair lift. Swisstopo As in [9], we use the positive and unlabeled F -score pre- sented in equation 6 to evaluate the performance and select 24 the best set of parameters for these two methods. Indeed, areas and geotagged with a GPS device. As stated above, for One-Class problems, the commonly used performance we want to learn which are the good places in term of land- measures, based on the ratio of false positives cannot be scape features: images taken in built areas tend to capture applied. the presence of a village or a touristic attraction rather than a natural landscape. The set of image locations retained contains 2683 points, which are then separated in the three 3. RESULTS subsets represented in figure 1. The first set, in blue, is used First, we will present the data and geographic features to train the methods; the second one, in green, is used to fit considered and how the image locations differ from the dis- the free parameters of the methods; the third one, in yellow, tribution of the random locations. Then, we conducted is used for the validation to compute independent statistics some experiment using different combination and different of the results. To generate those three sets, a grid with a amount of geographic features. Finally, we will present the 5km side is generated over the area. Then, each grid cell resulting probability maps. is randomly attributed to one of the subsets, in order to obtain spatially uncorrelated set of approximatively equal 3.1 Data size. Finally, we also select 10 sets of random locations with The experiments consider one of the political regions of the same number of locations as in the training set. These Switzerland located in the Swiss Alps: “Valais - Wallis”. The random sets will be used in the next sections to select sig- area is bounded by the Geneva Lake to the West and by the nificant geographic features and to compare the distribution highest summits of Switzerland. It encloses some of the most of the labeled and unlabeled data. touristic spots in Switzerland: Zermatt, the Matterhorn and the Aletsch glacier. The altitude gradient is important be- 3.2 Geographic features selection tween the lowest area on the Geneva lake shore (450m) and In order to measure the significance of these features, the the highest summit the “Dufourspitze” (4634m). The area unlabeled data and the picture location distributions are is a valley, whose bottom hosts small to mid-sized villages. compared. For each of the geographic feature chosen, their Climbing the flanks, the low altitudes are generally covered distribution diverge (tested by a Kolmogorov-Smirnov test with vineyards (500-700m); then forest, mountain villages with α = 1%). Despite their statistical differences, some and resorts are found in the range from 700 to 1400m; above features are more discriminative than others. The following 1400m pastures and slopes dedicated to ski give access to the list presents the geographic features selected and explains highest peaks, playground for the alpinists. how their value are different at true image locations. The Swiss Topographic Agency (Swisstopo) provides a DEM; among the available resolutions of 25m and 200m, • Altitude (Z): Since we are focusing on landscape we retained the first. Swisstopo also provides a TLM con- photography, few images are taken between 450 and taining vector layers of several territorial objects. For this 1500m. In contrary, the range between 1800 to 2200m, study, we selected the roads, other transportation facilities, where the ski slopes are found is very attractive. There forest and lake layers. The image locations are extracted is a small mode above 4000m representing pictures from the Flickr database and were provided by [25]. Those taken by alpinists at high altitude, figure 2 (a). images are filtered to keep only those located outside built • Curvature (Curv): Positive curvatures (convex area) are more represented; indeed the ridges are more at- tractive than the valley. • Slope: The flatter areas from 5% to 30%) are pre- ferred. • Visible Sky (Sky): This feature confirms the result observed with the curvature: cells with a high ratio of visible sky (>90%) are more often chosen. • Distance to nearest lake (DLake): People tend to take more pictures in a radius of 200m around the lakes. • Distance to nearest road (DRoad) and presence of roads (BRoad): The cells close to roads and paths are more active. Approximately 70% of the pictures are taken in a cell containing a road or a path, figure 2 (b). Figure 1: DEM hillshade rendering of the area under • Distance to the nearest forest (DFor): The ran- study overlaid by the Flickr image locations used: dom and image distributions are very similar. points colors depict whether the image is considered during training (blue), testing (green) or validation • Distance to the nearest cable car (DLift): Half (yellow). of the pictures are taken within a range of 1500m sur- rounding a lift. 25   !        (a)   (b) Figure 2: Distribution of the random and real locations for (a) Altitude, (b) Distance to the nearest road 3.3 Numerical results pictures aren’t taken from the ground (but from the lift) and In table 2, we report 10 experiments, obtained by consid- thus these locations aren’t related to the geographic features. ering different combination of geographic feature for GOCD. This also explains why the “DLift” geographic feature is not In this table only the best experiments are presented (C in the set of the best geographic features. Another set of lo- >= 20). Thus, the less-significant geographic features are cations, on the south of the “Breithorn” are badly classified less represented (DLift, DFor). The best combination of ge- (arrow B on the map in figure 3 (a)). They are shot on the ographic features is altitude, slope, visible sky, curvature, way to this peak which is one of the most easily accessible distance to the nearest road and the binary roads (Exp. 9 4000 summit in the Alps. However, the path to the summit in Table 2). The superiority of this experiment is observed is not in the roads / paths database. From this map, we for the two methods proposed but the KDE(KPCA) has the can understand the geographic features related to the image best result on the independent validation set, corresponding locations. First, the mountain paths are easily recognizable to a recall of 0.85. The KDE in the original space obtains in the figure. Indeed, they are always more attractive than slightly lower performance with a recall at 0.78. Thereby, the other locations. However, the paths on steep slopes are it appears that the KDE and KDE(KPCA) have a similar less probable than the other ones. Then, the ridges (Gorner- behavior for the combination of few significant geographic grat), passes and summits (Matterhorn) are more attractive feature (Exp. 1-4, 6, 7). However, if more features (Exp. 8 - than other areas. 10) or binary features (Exp. 5) are added, the KDE(KPCA) At the scale of the whole area, as seen in figure 4, it is in- is more suitable to describe the data relations between geo- teresting to note that the method is able to extract different graphic features, resulting to better results. Both KDE and behaviours. For instance, the paths are expected to have a KDE(KPCA) are more suited than KPCA and OCSVM for large probability. By combining the geographic features, it this problem. By inserting the unlabeled data in the clas- appears that it is true at medium altitude. However, in the sification process, we ensure that less unlabeled data are valley, where the roads have more traffic, locations between classified positive and thus less data in the independent val- the roads are preferred. Moreover, at higher altitude, where idation set are misclassified. ski slopes are found, people are more disposed to move away from the path. It is intuitive that the mountain peaks are 3.4 Evaluation on one of the most attractive good places to shot pictures. The strength of our method area in the Alps: Zermatt. is to rank the peaks according to their altitude and shape. On the map in figure 3 (a), probabilities for the Zermatt Finally, the less attractive locations are the steep slopes, bot- area are presented. The filled dots are the training loca- tom of deep valleys and flat areas at high altitude (glacier). tions, the triangles are the testing locations while the empty Indeed, these regions are hardly accessible. circles are the validation locations. Image locations are su- perposed to the results of KDE(KPCA) and correspond to the estimate of p(Y=1|z) at each location. A misclassifica- 4. CONCLUSION tion corresponds to a circle that would be located on an area The GPS measured positions of shared images are more of low probability (blue). The validation set is very specific accurate than locations provided by the user. In this paper, in this area: the pictures on the south are above 4000m we propose to use the GPS coordinates of a set of land- and accessible via a cable car (arrow A on the map in fig- scape pictures to train a classifier of likely and unelikely ure 3 (a)). This area is also a skiing region during winter image locations. Every map location is described with ge- and summer. First, some of the misclassified data are found ographic features extracted from a DEM and a TLM. The along the cable car and are easily explicable. Indeed, these method proposed projects the geographic features in a space 26 Table 2: This table summarizes the results of our experiments. C is the score proposed in equation 7 to evaluate our methods: KDE and KDE(KPCA) , Fpu is the score from equation 6 used to evaluate KPCA and OCSVM. For the method using KPCA the number of Principal Component Npc used is shown. nu is the ratio of outlier in OCSVM. R(V ) is the ratio of validation data correctly classified, this ratio is used to evaluate the methods. In bold, the best performing algorithm per experiment. Underlined the three best experiments over all tests. Geographic Proposed Competing Exp. features KDE KDE(KPCA) KPCA OCSVM considered C R(V) Npc C R(V) Npc Fpu R(V) Fpu ν R(V) 1 Z, slope, Sky, 33.3 0.83 4 34.5 0.83 3 1.94 0.45 1.66 0.4 0.41 DRoad 2 Z, slope, Sky, 19.2 0.73 4 24.86 0.72 5 2.43 0.4 1.19 0.45 0.48 Broad 3 Z, slope, Sky, 20.43 0.83 7 35.6 0.86 6 1.59 0.57 1.23 0.35 0.52 DRoad, Curv 4 Z, slope, Sky, 27.23 0.78 5 26.95 0.78 6 2.10 0.01 1.3 0.35 0.42 DRoad, DLake 5 Z, slope, Sky, 29.6 0.78 7 46.27 0.84 4 3.12 0.45 1.89 0.45 0.37 DRoad, BRoad 6 Z, slope, Sky, 18.8 0.69 4 25.7 0.55 4 1.52 0.55 1.18 0.25 0.64 DRoad, DFor 7 Z, slope, Sky, 20.7 0.75 5 23.54 0.7 2 2.11 0.42 1.21 0.2 0.59 DRoad, DLift 8 Z, slope, Sky, 21.85 0.79 7 34.44 0.86 6 1.52 0.53 1.16 0.25 0.63 DRoad, Curv, DLake 9 Z, slope, Sky, 35.26 0.78 7 51.74 0.85 5 2.31 0.55 1.44 0.35 0.52 DRoad, Curv, BRoad 10 Z, slope, Sky, 22.31 0.73 8 31.72 0.81 2 2.39 0.43 1.2 0.3 0.51 DRoad, Curv, Broad, DLake of higher dimension using a KPCA. Then, the spatial prob- the standard ones (OCSVM, KPCA) is mainly due to the ability density function of the picture locations and random insertion of the unlabeled data in the classification process. locations are estimated with a density function (KDE). The Our method is easy to implement and shows good results relation between them is used to compute the probability of without fine tuning. each map cell to be a likely location given the geographic We proved that the locations of a map are not equiprob- features. The recall on the independent validation set sur- able relative to landscape image locations. By describing passes the KPCA and OCVM classification in their regular each map location with geographic features, one can extract implementation. the most probable regions. The generated map differs from This preliminary study could be refined in several ways. the density maps based on the Northing and Easting of the First, the geographic features computed were selected from a picture locations in two ways: firstly, using the geographic priori expected correlations. However, other geographic fea- features, probabilities are also computed for region with- tures could also be correlated to the image locations (orien- out image locations. Second, by taking into account the tation, DEM based features at finer or coarser scales, rivers, locations accessibility, the computed probabilities are more cable car departure and arrival stations only etc.). Second, realistic. The force of our method is to learn attractive com- our studies focus on a small area in the Alps, a similar ap- binations of geographic features with the density estimation. proach could be applied to the entire Alps. Indeed, a bigger Indeed, the relation between a geographic feature and an the amount of picture locations would refine our results and in- picture location are intuitive (eg. paths are more probable, crease their robustness and generalization power at a larger convex area are preferred...), but their combinations is more scale. Third, in this work we avoided the KDE bandwidth powerful to classify correctly picture locations. setting using the Scott’s rule. A more appropriate band- Currently, there is a huge interest in computer vision for width could improve the results. Finally, the improvement the pose estimation of shared images at a local or worldwide between the proposed method (KDE and KDE(KPCA)) and scale. Our results show that it is possible to use geographic 27 0 0.5 1 (a) (b) Figure 3: Resulting maps: (a) Image locations are superimposed to the topographic map. Blue dots compose the training set, green triangles are the testing set, yellow circles are the validation set. ”A” shows the cable car; ”B” shows the way to the Breithorn. (b) Image locations are superimposed to probability map. Colour code is linked to the probability for a cell to be a picture location. Cells close to zero are similar to the random set, cells close to one are similar to the training set. Filled dots compose the training set, triangles are the testing set, empty circles are the validation set. features as a prior knowledge to either discredit some un- [6] C. Elkan and K. Noto. Learning classifiers from only likely poses or to promote the more probable ones. positive and unlabeled data. In International Conference on Knowledge Discovery and Data Mining, 5. ACKNOWLEDGMENTS 2008. [7] F. Girardin, F. Calabrese, F. Fiore, C. Ratti, and This work has been partly supported by the Swiss Na- J. Blat. Digital footprinting: Uncovering tourists with tional Science Foundation (grant PZ00P2-136827). user-generated content. Pervasive Computing, IEEE, 7(4):36–43, 2008. 6. REFERENCES [8] C. Grothe and J. Schaab. Automated footprint [1] G. Baatz, O. Saurer, K. Köser, and M. Pollefeys. Large generation from geotags with kernel density scale visual geo-localization of images in mountainous estimation and support vector machines. Spatial terrain. In ECCV, pages 517–530. Springer, 2012. Cognition & Computation, 9(3):195–211, 2009. [2] L. Baboud, M. Cadı́k, E. Eisemann, and H.-P. Seidel. [9] Q. Guo, W. Li, Y. Liu, and D. Tong. Predicting Automatic photo-to-terrain alignment for the potential distributions of geographic events using annotation of mountain pictures. In IEEE Conference one-class data: concepts and methods. International on Computer Vision and Pattern Recognition Journal of Geographical Information Science, (CVPR), pages 41–48. IEEE, 2011. 25(10):1697–1715, 2011. [3] C. Bozzini, M. Conedera, and P. Krebs. A new tool for [10] H. Hoffmann. Kernel PCA for novelty detection. obtaining cartographic georeferenced data from single Pattern Recognition, 40(3):863–874, 2007. oblique photos. Proceedings of the 23rd International [11] N. Jacobs, S. Satkin, N. Roman, R. Speyer, and CIPA Symposium, 2011. R. Pless. Geolocating static cameras. In International [4] P. Chippendale, M. Zanin, and C. Andreatta. Spatial Conference on Computer Vision, pages 1–6. IEEE, and temporal attractiveness analysis through 2007. geo-referenced photo alignment. In International [12] E. T. Jaynes. Information theory and statistical Geoscience and Remote Sensing Symposium mechanics. Physical review, 106(4):620, 1957. (IGARSS), volume 2, pages 1116–1119, 2008. [13] P. S. Kane. Assessing landscape attractiveness: a [5] F. de Morsier, D. Tuia, M. Borgeaud, V. Gass, and comparative test of two new methods. Applied J.-P. Thiran. Semi-supervised novelty detection using Geography, 1(2):77 – 96, 1981. SVM entire regularization path. IEEE Trans. Geosci. [14] J. Muñoz-Marı́, F. Bovolo, L. Gómez-Chova, Remote Sens., 51(4):1939–1950, 2013. 28 Figure 4: Resulting maps for the whole area. The case study area and Zermatt are highlighted on the Switzerland map. The colour code is the same as the previous map. L. Bruzzone, and G. Camps-Valls. Semisupervised [20] B. Schölkopf and A. Smola. Learning with Kernels. one-class support vector machines for classification of MIT press, Cambridge (MA), 2002. remote sensing data. IEEE Trans. Geosci. Remote [21] B. Schölkopf, R. C. Williamson, A. J. Smola, Sens., 48(8):3188–3197, 2010. J. Shawe-Taylor, and J. C. Platt. Support vector [15] E. Parzen. On estimation of a probability density method for novelty detection. In NIPS, volume 12, function and mode. The Annals of Mathematical pages 582–588, 1999. Statistics, 33(3):1065–1076, 09 1962. [22] D. W. Scott Satk. On optimal and data-based [16] A. Popescu and G. Grefenstette. Deducing trip related histograms. Biometrika, 66(3):605–610, 1979. information from flickr. In International Conference [23] Y. Sun, H. Fan, M. Bakillah, and A. Zipf. Road-based on World Wide Web, WWW ’09, pages 1183–1184, travel recommendation using geo-tagged images. New York, NY, USA, 2009. ACM. Computers, Environment and Urban Systems, (0), [17] T. Produit, D. Tuia, F. Golay, and C. Strecha. Pose 2013. estimation of landscape images using DEM and [24] D. M. Tax and R. P. Duin. Support vector data orthophotos. In International Conference on Computer description. Machine learning, 54(1):45–66, 2004. Vision in Remote Sensing, pages 209–214. IEEE, 2012. [25] T. Wider, D. Palacio, and R. S. Purves. [18] T. Produit, D. Tuia, C. Strecha, and F. Golay. An Georeferencing images using tags: application with open tool to register landscape oblique images and flickr. In AGILE International Conference on generate their synthetic model. In Open Source Geographic Information Science, 2013. Geospatial Research and Education Symposium, 2013. [26] D.-Y. Yeung and C. Chow. Parzen-window network [19] M. Rosenblatt. Remarks on some nonparametric intrusion detectors. In International Conference on estimates of a density function. The Annals of Pattern Recognition, volume 4, pages 385–388. IEEE, Mathematical Statistics, 27(3):832–837, 09 1956. 2002. 29