Computer Science Technical Report Archive

USC Computer Science Technical Reports, no. 912 (2009)

(USC DC Other)

USC Computer Science Technical Reports, no. 912 (2009)

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Vector Model in Support of Versatile Georeferenced Video Search Seon Ho Kim University of the District of Columbia Washington, DC 20008 skim@udc.edu Sakire Arslan Ay University of Southern California Los Angeles, CA 90089 arslan@usc.edu Byunggu Yu University of the District of Columbia Washington, DC 20008 byu@udc.edu Roger Zimmermann National University of Singapore Singapore 117590 rogerz@comp.nus.edu.sg ABSTRACT Increasinglygeographicpropertiesarebeingassociatedwith videos, especially those captured from mobile cameras. The meta data from camera-attached sensors can be used to model the coverage area of the scene as a spatial object such that videos can be organized, indexed and searched based on their ¯eld of views (FOV). The most accurate representation of an FOV is through the geometric shape of a circular sector. However, spatial search and indexing methods are traditionally optimized for rectilinear shapes because of their simplicity. Established methods often use anapproximationshape, suchasaminimumboundingrect- angle (MBR), to e±ciently ¯lter a large archive for possi- bly matching candidates. A second, re¯nement step is then applied to perform the time-consuming, precise matching function. MBR estimation has been successful for general spatial overlap queries, however it provides limited °exibil- ity for georeferenced video search. In this study we propose a novel vector-based model for FOV estimation which pro- vides a more versatile basis for georeferenced video search while providing competitive performance for the ¯lter step. Wedemonstratehowthevectormodelcanprovideauni¯ed methodtoperformtraditionaloverlapquerieswhilealsoen- ablingsearchesthat,forexample,concentrateonthevicinity ofthecamera'spositionorharnessitsviewdirection. Tothe bestofourknowledgenocomparabletechniqueexiststoday. Categories and Subject Descriptors H.2.4[DatabaseManagement]: Systems|Multimediadata- bases; H.2.4 [Database Management]: Systems|Query processing;C.4[PerformanceofSystems]: Modelingtech- niques General Terms Algorithms, Measurement, Performance Keywords Video search, georeferencing, meta-data, GPS USC Computer Science Technical Report. . 1. INTRODUCTION Advances in sensor technologies allow video clips to be tagged withgeographic properties, such as camera locations from GPS and camera directions from digital compasses, while being collected. Importantly, such meta-data can be attached to the video streams automatically, hence allowing for the consistent annotation of large amounts of collected video contents and thus enabling various criteria for versa- tile video search. The captured geographic meta-data have a signi¯cant potential to aid in the indexing and searching of georeferenced video data at the high semantic level pre- ferred by humans. However, there has been little research on utilizing such meta-data for the systematic indexing and searching of video data. Some video data are naturally tied to geographic loca- tions. For example, video streams from tra±c monitoring may not have much meaning without their associated lo- cation information. Thus, associated applications typically let a user specify location information to retrieve the traf- ¯c video related to a point or region. In anticipation of future applications, more and more still images are auto- matically tagged with geographic data as high-end devices are equipped with various sensors. Example cameras in- clude the Sony GPS-CS1, the Ricoh 500SE, the Jobo Photo GPS,theNikonD90withGPS;theRicohSE-3andtheSol- meta DP with GPS plus compass. Many smartphones are now equipped with a camera, GPS and accelerometer. We expect that it will only be a matter of time until many cam- corders will include such sensors. The resulting fused video plus sensor data streams can provide an e®ective means to index and search videos, especially for large archives that handle an extensive amount of video data. Oneclassofvideosearchtechniqueshasnaturallyfocused on identifying objects within the captured content through sophisticated extraction methods at the signal level or min- ingassociatedtextualmeta-datadescription. Evennow,the geo-tagging data are mostly used for organizing or group- ing images based on location information in a simple and straightforward way. Furthermore, most videos captured are not panoramic and as a result the viewing direction becomes very important for human perception, and conse- quently for video searching. GPS data only identi¯es object locations and therefore it is imperative to investigate the natural concepts of a human viewing direction and a view point. For example, we may be interested in videos that show a building only from a speci¯c angle. The question arises whether a video database search can accommodate such human friendly concepts, i.e., whether it is possible to index the video data based on the human viewable space andthereforetoenabletheretrievalofmoremeaningfuland recognizable scene results for user queries. The collection and fusion of multiple sensor streams such as the camera location, the ¯eld-of-view (FOV), the direc- tion, etc., can provide a comprehensive model of the view- able scene. For example, one may model the viewable scene area (i.e., FOV) of video segments or frames using a pie- shaped geometric contour, i.e., a pure spatial object, thus transforming the video search problem into a spatial data selectionproblem. Theobjectivethenistoindexthespatial objects and to search videos based on the geographic prop- erties of videos. Beyond the importance of the geographic information where a video is taken, there are other obvi- ous advantages in exploiting the spatial properties of video because the operation of a camera is fundamentally related to geometry. When a user wants to ¯nd images of an ob- ject captured from a certain viewpoint and from a certain distance, these semantics can be interpreted as geometric relations between the camera and the object such as the Euclidean distance between them and the directional vector from the camera to the object. Thus, more meaningful and recognizableresultsmaybeachievedbyusingspatialqueries on georeferenced videos. In this study we propose a new vector-based approxima- tion model for e±cient indexing and searching of georef- erenced video based on an FOV model. The FOV model represents a viewable scene of images as a circular sector using the tagged camera location, the direction, the angle, and the maximum viewable distance as illustrated in Fig- ure 2. As we will show, this model provides a more versatile basisforgeoreferencedvideosearchwhileprovidingcompet- itive performance. We demonstrate how the vector model can provide a uni¯ed method to perform traditional overlap queries while also enabling searches that, for example, con- centrateonthevicinityofthecamera'sposition ortake into account its view direction. To the best of our knowledge no comparable technique exists today. 2. RELATED WORK Associating GPS coordinates with digital media (images andvideos)hasbecomeanactiveareaofresearch[15]. There has been signi¯cant research on organizing and browsing personal photos according to location and time. Toyama et al. [18] introduced a meta-data powered image search and builtadatabase,alsoknownasWorldWideMediaeXchange (WWMX), which indexes photographs using location coor- dinates (latitude/longitude) and time. A number of addi- tional techniques in this direction have been proposed [12, 14]. There are also several commercial web sites [1, 3] that allowtheuploadandnavigationofgeoreferencedphotos. All these techniques use only the camera geo-coordinates as the reference location in describing images. We instead rely on the ¯eld-of-view of the camera to describe the scene. More related to our work, Ephstein et al. [6] proposed to relate images with their view frustum (viewable scene) and used a scene-centric ranking to generate a hierarchical organiza- tion of images. Several additional methods are proposed for organizing[16,9]andbrowsing[7,17]imagesbasedoncam- era location, direction and additional meta-data. Although these research work are similar to ours in using the cam- era ¯eld-of-view to describe the viewable scene, their main contribution is on image browsing and grouping of similar Test on simple approximations (n1 objects) Test on exact geometry (n2 objects) negatives positives (n2) false positives true positives (n3) Filter Step Refinement Step All objects (n1) Figure 1: Illustration of ¯lter-re¯nement steps. images together. Some approaches [17, 10] use location and othermeta-data,aswellastagsassociatedwithimages,and the images' visual features to generate representative candi- dates within image clusters. Geo-location is often used as a ¯ltering step. Some techniques [6, 16] solely use camera location and orientation in retrieving the\typical views"of important objects. However then the emphasis is on the segmentation of image scenes and organizing photos based on image scene similarity. Our work describes a more broad scenariothatconsidersmobilecamerascapturinggeo-tagged videos and the associated view frustum, which is dynami- cally changing over time. There exist only a few systems that associate videos with their corresponding geo-location. Liu et al. [11] presented a sensor enhanced video annotation system (referred to as SEVA)whichenablessearchingvideosfortheappearanceof particular objects. SEVA serves as a good example to show howasensorrich,controlledenvironmentcansupportinter- esting applications. However it does not propose a broadly applicable approach to geo-spatially annotate videos for ef- fective video search. A recent study [2] investigated these issues and proposed a viewable scene model to describe the video content but did not address the search issues. 3. MODELING FOV USING VECTOR 3.1 Motivation When a large collection of videos is stored in a database, the cost of processing spatial queries may be signi¯cant be- cause of the computational complexity of the operations in- volved. Therefore, such queries are typically executed in two steps: a ¯lter step followed by a re¯nement step [13, 5] (Figure 1). The idea behind the ¯lter step is to approxi- mate the large number complex spatial shapes (n1 objects in Figure 1) with simpler outlines (e.g., a minimum bound- ing rectangle, MBR [4]) so that a large number of unrelated objects can be dismissed very quickly based on their simpli- ¯ed shapes. The resulting candidate set (n2 objects) is then further processed during the re¯nement step to determine the exact results (n3 objects) based on the exact geometric shapes. The rationale of the two step process is that the ¯l- ter step is computationally far cheaper than the re¯nement step due to the simple approximations. Overall, the cost of spatial queries is determined by the e±ciency of the ¯lter V1 x y VX V1Y VY V1X V1X V1Y Figure 2: FOV representation in di®erent spaces. step (many objects, but simple shapes) and the complexity of the re¯nement step (few objects with complex shapes). Additionally, in video search applications, the re¯nement step can be very expensive due to the nature of the pro- cessing. Depending on the application, various computer vision and content-based extraction techniques may be ap- plied before presenting the search results. For example, some occlusions may need to be detected based on local ge- ographic information such as the location and size of build- ings. Some speci¯c shapes or colors of objects might be analyzed for more accurate results, or the quality of images suchasbrightnessandfocusmaybeconsideredindetermin- ingtherelevancerankingofresults. Suchextraprocessingis in general performed during re¯nement on a per frame ba- sis, therefore signi¯cantly increases the time and execution cost of the re¯nement step. It is thus critical to minimize the amount of re¯nement processing for large scale video searches. This, in turn, motivates the use of e®ective and e±cient ¯ltering algorithms which minimize the number of frames that need to be considered in the re¯nement step. In traditional spatial data processing, MBR approxima- tions are very e®ective for the ¯lter step. However, with a bounding rectangle some key properties that are useful in video search applications may be lost. For example, MBRs retain no notion of directionality. This study advocates a new vector approximation that provides similar e±ciency and low processing cost as MBR-based methods, but ad- ditionally provides better support for the type of searches that a video database may encounter. Thus, the main focus of the paper is to provide a novel ¯lter step called the vec- tor model as a more e±cient and e®ective ¯lter step for the large scale georeferenced video search applications and to provide its comparison to the conventional ¯lter step using MBRs. The identical re¯nement step will be assumed for a fair comparison between the vector and MBR model. Inthefollowingsectionswewillintroduceourvectormodel and illustrate that it is both competitive with MBR-based methods where applicable, but also extends to cases that MBRs cannot handle. 3.2 Vector Model A camera positioned at a given point p in geo-space cap- tures a scene whose covered area is referred to as camera ¯eld-of-view(FOV,alsocalledaviewablescene). Themeta- data related to the geographic properties of a camera and its captured scenes are as follows: 1) the camera position p is the latitude, longitude coordinates read from a posi- tioning device (e.g., GPS), 2) the camera direction ® is obtained based on the orientation angle (0 ± · ® < 360 ± ) provided by a digital compass, 3) the maximum visible dis- tance from p is R at which objects in the image can be recognized by observers [2] { since no camera can capture meaningfulimagesataninde¯nitedistance,Risboundedby M which is the maximum distance set by an application {, and 4) the camera view angle µ describes the angular extent of the scene imaged by the camera. The angle µ is calcu- lated based on the camera and lens properties for the cur- rent zoom level [8]. The above geo-properties are captured fromasensor-equippedcamerawhilevideoisrecorded. Note that some commercial cameras already are expected to be equipped with those sensors in the very near future. Based on the availability of the sensor input data, the FOV of a video frame forms an area of circular sector shape (or pie-slice shape) in 2D geo-space as shown in Figure 2. Then, an FOV can be represented as a tuple <T;p;µ;V>, with T as the real time when the frame was captured, a positionp,anangleµ,andacentervectorV.Themagnitude ofVistheviewabledistancefromp,i.e.,Randthedirection of V is ®. For indexing purposes, we propose a vector estimation model that represents an FOV using only the camera posi- tion p and the center vector V. When we project the FOV onto the x and y axis, a point p is divided into px and py, and V is divided into V X and V Y along the x and y axis, respectively. Then, an FOV denoted by a point and vector can be represented by a quadruple ; this can be interpreted as a point in four dimensional space. In mathematics, space transformation is an approach to simplifythestudyofmultidimensionalproblemsbyreducing them to lower dimensions or by converting them into some othermultidimensionalspace. Usingaspacetransformation, an FOV <px;py;VX;VY> can be divided and represented in two 2D subspaces, i.e., p x ¡V X and p y ¡V Y . Then, an FOV can be represented as two points, each in its own 2D space. Forexample,Figure2showsthemappingbetweenan FOV represented by p1 andV1 in geo-space and two points in two transformed spaces without loss of information. To de¯ne the vector direction, let any vector heading towards the right (East in the northern hemisphere) on the x axis have a positive V X value, and a negative V X value for the other direction (West). Similarly, any vector heading up (North) on y axis has a positive VY value, and a negative V Y valuefortheotherdirection(South). Usingtheproposed model, any single FOV can be represented as a point in a p¡V space. As a result, the problem of searching for FOV areas in the original space can be converted to the problem of ¯nding FOV points in the transformed subspace. Note that the actual FOV is an area represented by a circular sector, so representing an area using a single vector is incomplete. More precisely, the FOV can be considered x y q q q - M q + M q q -M q +M M -M M px v v py v v v v & v 0 0 q q q -M q +M M -M Figure 3: Illustration of ¯lter step in point query processing. x y q q q - M q + M q q -M q +M M -M px v v v py v v v v v 0 0 q v v q q -M q +M M -M Figure 4: Example of ¯ltering in point query. as a collection of vectors starting from p to all the points on the arc. To simplify the discussion for now we use only one center vector to represent an area as described above. We will relax this simplifying assumption in Section 5. 4. QUERY PROCESSING When we represent video content as a series of FOVs whichhaveaspeci¯cshape, FOVscanbeconsideredasspa- tialobjects. Theproblemofvideosearchisthentransformed into ¯nding spatial objects in a database. This section de- scribes how the ¯lter step can be performed by using the proposed vector model for some typical spatial query types. 4.1 Point Query Theassumedqueryis,\Foragivenquerypoint q <x;y > in 2D geo-space, ¯nd all video frames that overlap with q." The¯lterstepcanbeperformedinp¡V spacebyidentifying all possible points of FOVs that have a potential to overlap with the query point. Recall that the maximum magnitude of any vector is lim- ited to M, and hence any vector outside of a circle centered at the query point q with a radius M cannot reach q in geo- space; see Figure 3 for an illustration. Only vectors starting inside the circle (including the circumference of the circle) haveapossibilitytocrossormeetq. Theobjectiveofthe¯l- ter step is to search for such vectors in the p¡V spaces, not in the x¡y space. Because a query point is not a vector, it is mapped only to the p axis. First, let us consider only the x components of all vectors. In p x ¡V X space, the possible vectors that can cross (or touch) q x should be in the range [qx¡M;qx +M]. That is, any vector at px is ¯rst ¯ltered out if jpx¡qxj > M. Next, even though a vector is within the circle, it cannot reach q x if its magnitude is too small. Thus,jp x ¡q x j·jV X j must be satis¯ed for V X to reach q x . At the same time the vector direction should be towards q x . For example, when px > qx, any vector with a positive VX value cannot meet qx. Hence, in p¡V spaces as shown in Figure 3, all points (i.e., all vectors) outside of the shaded isosceles right triangle areas will be excluded in the ¯lter step. For example, vector V 1 in geo-space is represented as a point v1 in p¡V space. Now consider all vectors starting from a point on the circumference of the circle towards the center with the maximum magnitude M. All such vectors movingfromV 1 toV 4 inaclockwisedirectionmaptothedi- agonal line starting from v 1 to v 4 in p¡V space. The same can be observed for the y components of vectors, i.e., the same shape appears in p y ¡V Y space. The resulting vectors from the ¯lter step should be included in the shaded areas of both p x ¡V X and p y ¡V Y space. Formally, a vector at p that satis¯es the following conditions can be selected in the ¯lter step: jp¡qj·M p x ¡q x ·¡V X if p x >q x p y ¡q y ·¡V Y if p y >q y qx¡px·VX if qx >px qx¡py ·VY if qy >py any V X if q x =p x any V Y if q y =p y (1) Figure 4 shows ¯ve examples of FOVs and their mapping between x¡y space and p¡V spaces. The starting points of all ¯ve vectors are within the circle. However, not all of them pass the ¯lter step. The starting points of three vec- tors, V 1 , V 2 , and V 4 , are located inside the circle but their vector direction and/or magnitude do not meet the neces- sary conditions so they are ¯ltered out. For example, V1x is heading in the opposite direction even though its magni- tude is large enough. Thus, v 1x is outside of the triangle x y q q q - M q + M q q -M q +M M -M M px py 0 0 q q r q +r q +r q -r q -r q -M q +M M -M Figure 5: Illustration of the ¯lter step in point query with bounded distance r. x y q q q - M q q +M -M orth px q β q -Msinβ q q -M M py q -Msinβ q -Mcosβ M Msinβ Mcosβ q -Mcosβ Figure 6: Illustration of ¯lter step in directional point query with angle ¯. shape search space. Similarly, V4x is heading in the wrong direction so v 4x is outside of the search space. V 5 is directly headingtowardsq andbothV 5x andV 5y havealargeenough magnitudetoreach q. Thus, bothv 5x andv 5y areinsidethe search space, which means the vector should be included in the ¯lter result. V3 is considered a false positive in the ¯lter result because it satis¯es the conditions but actually does not cover q. It will be pruned out in the re¯nement step. 4.2 Point Query with Bounded Distance Unlike with a general spatial query, video search may en- force application speci¯c search parameters. For example, onemightwanttoretrieveonlyframeswhereacertainsmall object at a speci¯c location appears within a video scene, but with a given minimum size for better visual perception. Usually, when the camera is close to the query object, the object appears larger in the frame. Thus, we can devise a searchwitharangerestrictionforthedistanceofthecamera locations from the query point such as\For a given query point q <x;y > in 2D geo-space, ¯nd all video frames that overlap with q and that were taken within the distance r fromq." Becauseofthedistancerequirement r, theposition of the camera in an FOV cannot be located outside of the circle centered at q with radius r, where r < M. Thus, the search space can be reduced as shown in Figure 5. 4.3 Directional Point Query The camera view direction can be an important factor for the image perception by an observer. Consider the case where a video search application would like to exploit the collectedcameradirectionsforquerying. Anexamplesearch is,\For a given query point q <x;y > in geo-space, ¯nd all video frames taken with the camera pointing in the North- west direction and overlapping with q." The view direction canbede¯nedasalineofsightfromthecameratothequery point (i.e., an object or place pictured in the frame). The lineofsightcanbede¯nedusinganangleatthecameraloca- tionsimilartothecameradirection®. Notethatthecamera orientationis always pointingto the center of anFOV scene whiletheviewdirectioncanpointtoanylocationsorobjects within the scene. A digital compass mounted on a camera willreportthecameradirectionprimarilyusingbearings. A bearingisahorizontalanglemeasuredclockwisefromNorth (eithermagneticNorthortrueNorth)toaspeci¯cdirection. Whenweusebearingastheviewdirectionangle(say¯),the Northwest direction is equivalent to 315 degrees (Figure 6). An important observation is that all FOVs that cover the query point have their starting points along the same line of sight in order to point towards the requested direction. Thus, the ¯lter step needs to narrow the search to the vec- tors that satisfy the following conditions: 1) their starting points are on the line of sight, 2) their vector directions are heading towards q, and 3) their vector magnitudes are long enough to reach q. For a given view direction angle ¯, we can calculate the maximum possible displacement of a vector starting point from the query point. Because the largest magnitude of any vector is M, the maximum displacement between the query point and the starting point of any possible overlap- ping vector is ¡Msin¯ on the x axis and¡Mcos¯ on the y axis (note that the sign is naturally decided by ¯, e.g., sin315 ± = ¡0:71 and cos315 ± = 0:71). In other words, as showninFigure6,anyvectorstartingatapointgreaterthan q x +(¡Msin¯) on the x axis or less than q y +(¡Mcos¯) on the y axis cannot touch or cross the query point with the given angle ¯. Thus, the search area for such vectors can be reduced as illustrated in Figure 6. To meet the view direction request (say, 315 ± line of sight), no vector with a x y q q q - M q q +M -M orth px q β q -Msinβ q M py q -Msinβ M rsinβ r q -Mcosβ q -rcosβ q -rsinβ q -rsinβ Msinβ q -M q +Mcosβ q +rscosβ rcosβ Mcosβ Figure 7: Illustration of ¯lter step in directional point query with ¯ and r. x q x x −M x x +M x x -M x +M M px x -M 0 y y -M y +M +M -M py y y y y + y − Figure 8: Illustration of ¯lter step in range query. positive VX value can reach q. Therefore, in the ¯lter step the entire search space (i.e., the triangle shape) on the pos- itive V X side is excluded in the p x ¡V X space. Similarly, no vector with a negative V Y value can reach q, so the en- tire search space (the triangle shape) on the negative VY side is excluded in the px¡VY space. Next, the size of the remaining search space is reduced because the range of pos- sible V X and V Y values is now [0;Msin¯] and [0;Mcos¯], respectively. Using only a single speci¯c view direction value may not be practical in video search because a slight variation in viewdirectionsdoesnotsigni¯cantlyalterthehumanvisual perception. Therefore, it will be more meaningful when the queryisgivenwithacertainrangeofdirectionssuchas¯§², e.g, 315 ± §10 ± . The extension can be straightforward and it will increase the search area in the p¡V space. 4.4 Directional Point Query with Bounded Dis- tance This type of query is a hybrid of the previous types. For a very speci¯c search, the user might specify the query po- sition, the view direction from the camera, and the distance between the location of the query and the camera. An ex- ample query is, \For a given query point q < x;y > in geo-space, ¯nd all video frames heading in the Northwest direction, overlapping with q and taken within the distance r from q." The objective of this query is to ¯nd frames in whichsmallobjects(e.g.,a6meterhighstatue)atthequery point appear large in the viewable scenes. Another example query is,\For a given query point q < x;y > in geo-space, ¯ndallvideoframesheadingintheNorthwestdirectionthat overlapwithq andthatweretakenfartherthanthedistance r from q." Now the intention is to ¯nd frames where large objects (e.g., a 6 story tall building) at the query point ap- pear prominently in the frames. For the former case, the positionsofcamerasareboundedby¡rsin¯ fromq x onthe p x axis and ¡rcos¯ from q y on the p y axis, respectively. At the same time, the vector is bounded by rsin¯ on the VX axis and rcos¯ on the VY axis, respectively. Therefore, the grid patterned triangle area in Figure 7 represents the search space. For the latter case, the positions of cameras are bounded within [¡rsin¯;¡Msin¯] on the p x axis and [¡rcos¯;¡Mcos¯] on the p y axis, respectively. Further- more, the vector is bounded within [rsin¯;Msin¯] on the VX axis and [rcos¯;Mcos¯] on the VY axis, respectively. Therefore, the shaded triangle area in Figure 7 represents the search space. 4.5 Rectangular Range Query The assumed query is, \For a given rectangular query range in geo-space, ¯nd all the video frames that overlap withthisregion." Assumethattherectangularqueryregion q is a collection of points (the rectangular shaded area with a grid pattern in Figure 8). When we apply the same space transformation, all points in the query region can be repre- sented as a line interval on the px and py axes. First, when any vector's starting point falls inside the query region, the vectorclearlyoverlapswith q soit shouldbe includedin the result of the ¯lter step. Next, when we assume that any location along the perimeter of q is an independent query point as in Section 4.1, the starting points of vectors that can reach the query point is bounded by a circle with radius M. Drawing circles along all the points on the perimeter forms the shaded region in Figure 8. It follows that any vector with its starting point outside of the shaded region cannot reach any point in q. Only vectors starting inside the region have a possibility to cross q. The search area in the p¡ V spaces can be de¯ned as x y p δx δy F V C ϴ V L V R α px V py V V V V V Figure 9: Problem of single vector model in point query processing. x y q - M + M -M +M M -M F1 px py 0 0 F2 V1 V2 q +M δ δ q q -M M -M Figure 10: Overestimation constant ±. shown in Figure 8. Again, any vector in the resulting set should be found in both search areas in the p x ¡V X and p y ¡ V Y spaces. When p x and p y of a vector fall inside the mid-rectangles (grid pattern), p is inside q so the vector automatically overlaps with q regardless of its direction or magnitude. However, when p is located outside of q, the vector's direction and magnitude should be considered to determine the overlap. 5. IMPLEMENTATION Sofar,weassumedthatanFOVisrepresentedbyasingle center vector. However, in reality, an FOV is a collection of vectors with the following properties: 1) they all start from the same point, 2) they have the same magnitude jVj, and 3) they have di®erent directions to points along the arc of a circular sector. In this paper we de¯ne an FOV using <T;p;µ;V>,whereV isthecentervectorofaFOV(i.e., VC in Figure 9). V C consists of a compass bearing ® as direc- tion and the visible distance R as magnitude. When only a single vector VC represents the entire area of an FOV, there is a limitation in retrieving all the objects covered by the FOV. Because V CX and V CY are used to represent the FOV in p¡ V spaces as described in Section 3, this approach underestimates the coverage of the FOV. In Figure 9, the rectangle with the grid pattern represents the estimation of theFOVinthe¯lterstepusingVC. Onlyquerypointsinside the rectangle are selected during the ¯lter step. The black dots overlap with the actual FOV so they represent the true query results. The white dots overlap with the estimation of the FOV but they are not actually overlapping with the FOV. The single vector model cannot exclude these points during the ¯lter step, thus they become false positives. The problem is that the white rectangles are ¯ltered out even though they are actually inside the FOV. They are com- pletely missed during the search. Alternatively, one can use two vectors to represent an FOV, the leftmost and the rightmost vector (V L and V R ). Bothhavethesamemagnitudebutdi®erentdirections(their calculation from the collected data VC is straightforward). When we use V L and V R to estimate the FOV, the estima- tion area is extended by ± x and ± y along the x and y axis, respectively (grid rectangle plus shaded areas in Figure 9). Thisapproachcanencompasstheblackdots,thewhitedots, andthewhiterectangles, whichmeansthatitisnotmissing any query points within the FOV. However, the number of false positives may also increase due to the bigger estima- tion area. The triangular points in the ¯gure become false positives which are ¯ltered out in the single vector model. Notethatthetwo-vectormodelnowhasanidenticalestima- tion size compared with the MBR model. The more serious problemisthatthetwo-vectormodelmakestheuseof p¡V space as search region more complex because we cannot use asimplepointqueryinp¡V space. ThisisbecauseanFOV isrepresentedbyalineintervalboundedbythetwovectors, as seen in Figure 9. The exact boundary is related to the camera direction ®, the angle µ, and the visible distance R. This problem can be resolved when we introduce an over- estimation constant ± in de¯ning the search area in p¡V space. The overestimation constant is a generalization of errors in using a single vector model, i.e., ± x and ± y . As shown in Figure 10, a single vector V1 represents an FOV, F1. This vector covers the query point in the middle of an FOV and so it can be searched without any problem using thetriangularshapedsearchspaceasoriginallydescribedin Section4.1. However, the other vectorV2, representing F2, cannot be included in the search space. Because the query point is located at the leftmost corner of F2, v 2y covers q y but v 2x falls outside of the triangle. V2 is not considered as overlapping so F2 is missed. However, if the search space is extended by ± along the V axis (the parallelogram-shaped shaded area), v 2x becomes included in the search space and V2canbeselectedinthe¯lterstep. Notethat,inFigure10, ± is applied in one direction because the other direction al- ready reaches the maximal value of M. The next question is how to de¯ne the overestimation constant ±. The overestimation constant can be determined by the tolerable error between the magnitude of the center vector and the leftmost (or rightmost) vector as explained above. Assuming a regular camera (non-panoramic), the camera angle µ can be 180 ± in the worst case, which results in the maximum di®erence M. This maximum value, i.e., ± = M, signi¯cantly increases the search area in p¡V space. As µ becomes smaller, the extended search area decreases. The range of the overestimation constant is 0 · ± · M. However, note that normal camera lenses generally cover between 25 ± to 60 ± and wide angle lenses cover between 60 ± to 100 ± . Only ultra wide angle lenses capture up to 180 ± . Aninterestingobservationontheoverestimationconstant is that it can be an important parameter of georeferenced video search. First, the angle µ is related to the zoom level ofthecameraandthevisibledistanceR. Foracertainangle µ, the overestimation constant is limited to Msin(µ=2) for 100% coverage. In our experiments, the widest measured angle was 60 ± and the maximum visible distance was 259 meters. In this case the worst overestimation constant will be 259£sin(60=2) = 129:5 meters. Another important ob- servation in video search is that small objects which cannot be easily perceived by humans may be sometimes ignored even though they actually appear in FOVs. For example, if an object appears in the far left corner of an FOV and occupies only a very small portion of the frame, users may not be interested in such results. Moreover, if an object is located very far from the camera location (i.e., near the arc in our proposed model), it might be blocked by some nearer objects. Di®erent applications (or users) might require dif- ferentlevelsofaccuracyinsearchresults. Sotheoverestima- tion constant provides a tradeo® between the performance andtheaccuracyofvideosearch. Notethatasmalleroveres- timationconstant¯ndsFOVswherethequerypointappears inthecenterpartoftheframesande®ectivelydiscriminates againstotherframeswherethequerypointappearstowards the far left or far right side of the frames. 6. EXPERIMENTAL EV ALUATION 6.1 Data Collection and Methodology To collect real georeferenced video data, we have con- structed a prototype system which includes a camera, a 3D compass and a GPS receiver. We used the JVC JY-HD10U camera with a frame size of approximately one megapixel (1280£720 pixels at a data rate of 30 frames per second). It produces MPEG-2 HD video streams at a rate of slightly morethan20Mb/sandvideooutputisavailableinrealtime from the built-in FireWire (IEEE 1394) port. To obtain the orientationofthecamera,weemployedtheOS5000-USSolid StateTiltCompensated3AxisDigitalCompass,whichpro- vides precise tilt compensated headings with roll and pitch data. To acquire the camera location, the Pharos iGPS-500 GPS receiver has been used. A program was developed to acquire, process, and record the georeferences along with the MPEG-2 HD video streams. The system can process MPEG-2 video in real-time (without decoding the stream) and each video frame is associated with its viewable scene information. In all of our experiments, an FOV was con- structed every second, i.e., one FOV per 30 frames of video. We mounted the recording system setup on a vehicle and capturedvideodrivingalongstreetsatdi®erentspeeds(max. 25 MPH). During video capture, we frequently changed the camera view direction. The recorded videos covered a 5.5 kilometer by 4.6 kilometer region quite uniformly. However, forafewpopularlocationsweshotseveralvideos,eachview- ing the same location from di®erent directions. The total captured data includes 134 video clips, ranging from 60 to 240 seconds in duration. Each second, an FOV was col- lected, resulting in 10,652 FOVs in total. We generated 1,000 point queries which were randomly distributed within the region. Figure 11.(a) shows the distribution of the cam- era positions of 10,652 FOVs and the 1,000 query points. For each query, we searched the georeferenced meta-data to ¯nd the FOVs that overlap with that query. ForallexperimentsweconstructedalocalMySQLdatabase thatstoredalltheFOVmeta-dataandtheirapproximations (both MBRs and vectors). We used MySQL Server 5.1 in- stalled on a 2.33 GHz Intel Core2 Duo Windows PC. For each query type described in Section 4 with the MBR and thevectorapproximation,wecreatedaMySQLuserde¯ned function(UDF)tosearchthroughtheFOVsinthedatabase. We also implemented a UDF for the re¯nement step which returnstheactualoverlapinstancesbetweenaqueryandan FOV. We used the Universal Traverse Mercator coordinates for all comparisons. For the evaluation of the search results with di®erent ap- proaches, we computed the recall and precision metrics for the¯lterstep. Therecallisde¯nedasthenumberofoverlap- ping FOVs found in the ¯lter step by an approach over the actual number of overlapping FOVs. Note that the actual number of overlapping FOVs is obtained after the re¯ne- ment step from the exact geometric calculation using the circular sectors. The precision is de¯ned as the number of actually overlapping FOVs found over the total number of FOVs returned in the ¯lter step. 6.2 Comparison WesetthedistanceM tothemaximumviewabledistance amongallrecordedRvaluesofFOVs,soM equaled259me- ters. Thewidestcameraanglerecordedwas60 ± . Thus,inall experimentswesetthemaximumoverestimationconstantto sin30 ± £259, i.e., 0:5M. Point query: After executing 1,000 random point queries with 10,652 FOVs in the database, the number of actual overlap instances between a query point and an FOV was 17,203. Thisnumberwasobtainedandveri¯edbygeometric computation after the re¯nement step, i.e., it represents the ground truth. The point query results from the MySQL im- plementation are summarized in Table 1 and Figure 11.(b). TheMBRapproachreturned30,491potentiallyoverlapping FOVs in the ¯lter step and found all 17,203 actually over- lappingFOVsatthere¯nementstep. Thevectormodelwas applied with varying ± values. As expected, with the max- imum overestimation constant ± = 0:5M the vector model showed almost identical results to those of the MBR model (thesizeoftheapproximationisslightlybiggerthanwithan 5173140 5173640 5174140 5174640 5175140 5175640 5176140 5176640 5177140 497460 498460 499460 500460 501460 502460 camera position querypoint 0 5000 10000 15000 20000 25000 30000 35000 MBR 0.5M 0.4M 0.3M 0.2M 0.1M 0.0M FOVs returned FOVs matched (a) Camera positions and query points (b) Retrieved FOVs for point query with r =M 0.0 0.2 0.4 0.6 0.8 1.0 1.2 MBR 0.5M 0.4M 0.3M 0.2M 0.1M 0.0M PQ PQ-r50 PQ-r100 PQ-r150 PQ-r200 0.0 0.2 0.4 0.6 0.8 1.0 MBR 0.5M 0.4M 0.3M 0.2M 0.1M 0.0M PQ PQ-r50 PQ-r100 PQ-r150 PQ-r200 (c) Recall with varying r (d) Precision with varying r 0 500 1000 1500 2000 30491 MBR 0.5M 0.4M 0.3M 0.2M 0.1M 0.0M FOVs returned FOVs matched 0 10000 20000 30000 40000 50000 60000 70000 MBR 0.5M 0.4M 0.3M 0.2M 0.1M 0.0M FOVs returned FOVs matched (e) Retrieved FOVs for point query with r =50m (f) Retrieved FOVs for range query Figure 11: Summary of experimental results. MBR). However, when we decreased the value of ±, the vec- tormodelreturnedasmallernumberofactuallyoverlapping FOVsaswellasasmallernumberofpotentiallyoverlapping FOVs at the ¯lter step. This is because the vector model is discriminating more against overlapping objects at the side of scenes as the value of ± decreases. Figure 12 provides an example of how di®erent approaches perform the ¯lter step. The MBR for the 42 nd FOV of video 61 overlapped with 7 query points while only 6 points actually overlapped with the FOV. The vector model found di®erent numbers of query points as ± varied. As shown in Figure 12.(a), query pointsA,B,andGwerelocatedclosertothecentervectorof the FOV, so they were found in all approaches, even when ± =0:0M. However, D and F were very far from the center vector so they were only found when ± became larger. The vector model discriminates against objects at the side of an FOV as ± gets smaller. The vector model with a reduced ± found a smaller number of FOVs. The query points closer to the sides (i.e., those that may not be well perceived by humans) were e®ectively excluded. Overall, when ± grows the recall increases and the precision decreases. Wemeasuredthetimetoexecute1,000pointquerieswith MySQL using various approaches. The bottom row of Ta- ble 1 shows the total amount of time in seconds reported by MySQL. On average, the vector models took 14-19% more time than the MBR model. We did not use indexes in the search so the results re°ected the computational time for table scans. In reality, indexes such as B-trees or R-trees are used for a more e±cient spatial search for a larger set of dataandtheexecutiontimeofthe¯lterstepdependsonthe performance of the indexes. Note that the reported time is forthe¯lterstepsincethere¯nementstepwasimplemented as a separate program with the results from MySQL. The focus of this study is not on the speedup of the ¯lter step itself but on the overall query processing, and the e®ective- ness of the ¯lter step to support versatile search using the characteristics of video. Even though the execution time of the vector-based ¯lter step is a little longer than that of Table 1: Detailed results of point query. MBR Vector (with di®erent overestimation constant) 0.5M 0.4M 0.3M 0.2M 0.1M 0.0M FOVs returned 30491 32535 28302 23843 19268 14762 10360 FOVs actually matched 17203 17197 16620 15390 13686 11488 8493 Recall 1.00 1.00 0.966 0.895 0.796 0.668 0.494 Precision 0.564 0.529 0.587 0.645 0.710 0.778 0.820 Exec. time of 1000 queries (sec) 8.5 10.5 10.5 10.0 10.0 9.7 9.7 theMBR-basedone,thenumberofselectedobjectsfromthe vectormodelcanbefarsmallerthanthatoftheMBRmodel as shown in Tables 1, 2 and 3, which results in a signi¯cant speed up on the overall query processing by minimizing the workload of time-consuming re¯nement step. Point query with bounding distance r: Figure 11.(e) shows the results of point queries with a bounding distance r between the camera position and the query point. When r was 50 meters the number of matching FOVs for 1,000 querieswas649. Notethat50misapproximatelyone¯fthof the maximum viewable distance, which means that overlap- pingquerypointsshouldbecontainedin1/25oftheoriginal FOV size. Thus, the number of overlap instances is greatly reduced. The MBR model returned the same 30,491 FOVs but, for example, the vector model with ± =0:5M returned only1,908(a94%reduction)with1.0recall. As ± decreased the recall diminished as well and the precision increased. This trend is analogous to the one observed on the results of point queries without a bounding distance. We repeated the same experiments while varying r from 50m to 200m. The results all exhibited the same trend as shown in Fig- ures 11.(c) and 11.(d). Figure13illustratesthee®ectsofthervalueonthesearch. WhenwesearchedforquerypointF (thePizzaHutbuilding inthe scenes)without r (i.e., r =M), bothframes shownin Figures 13.(a) and (c) were returned because they contain the query point. However, the building appears very small (andisdi±culttoberecognizedbyhumans)inFigure13.(c) sinceitwaslocatedfarfromthecamera. Notethatthesame building is easily recognizable in (a) when the camera was closer to the object. We can e®ectively exclude (c) using an appropriate r value. Figures 13.(b) and (d) show the alternative satellite images of (a) and (c), respectively. Directional point query: Using the same 1,000 query points,wesearchedforallFOVsthatoverlapwiththequery points while varying the viewing direction from the camera to the query point. We used a §5 ± error margin with the viewing direction in all experiments. Table 2 shows the re- sultsofpointquerieswitha45 ± viewingdirection. TheMBR approach has no information about the direction in its es- timation so it resulted in the same number of 30,491 FOVs which must be processed in the re¯nement step. When the overestimation constant is not too small (± ¸ 0:3M), the vector model resulted in an approximately 90% reduction in the number of selected FOVs in the ¯lter step compared to the MBR method, while providing a recall value of over 0.9. Signi¯cantly { as shown in Figure 12 { the missing FOVs mostly contained query points at the far sides of the viewablescenearea. Fordi®erentviewingdirections,similar results were observed. Directional point query with bounding distance r: Table 3 shows the results of a very speci¯c point query case, i.e., one that considers both the viewing direction and Table2: Resultsofdirectionalpointquerywith 45 ± § 5 ± . The actual number of matched FOVs were 402. MBR Vector 0.5M 0.3M 0.1M FOVs returned 30491 3858 2972 720 FOVs actually matched 402 389 381 134 Recall 1.000 0.968 0.948 0.333 Precision 0.013 0.101 0.128 0.186 bounding distance. The vector model e®ectively excludes non-overlapping FOVs in the ¯lter step. For example, with a 45 ± viewing direction and r = 50m there were only 13 overlapping instances, which is a very small number with respect to the 1,000 queries and 10,652 FOVs. The vector model returned 374 FOVs, including the matched 13. Note that the MBR model returned 30,491 FOVs. We repeated thesameexperimentswhilevarying ± andobservedthatthe vector model provided the best balance between recall and precision with a value of ± =0:3M. Table 3: Results of directional point query with r. 45 ± §5 ± viewing direction and ± =0:3M. Vector r=50 r=100 r=150 r=200 FOVs returned 374 1006 2124 2972 FOVs actually matched 13 93 151 264 Recall 1.000 1.000 1.000 1.000 Precision 0.035 0.092 0.071 0.089 Range query: We generated 1,000 random queries with an identical query region size of 100m by 100m, but di®er- ent locations. For each range query, we checked the overlap between the query area and the FOVs. Figure 11.(f) sum- marizes the results, which show a similar trend as observed with point queries. The vector model with ± = 0:5M pro- vided almost perfect recall, namely 0.998, with a slightly higher number of FOVs returned in the ¯lter step. As ± diminishes the recall decreases and the precision increases. The chances for overlap between a given query range and any FOV increases as the size of the query range becomes larger. When we increased the query size to 300m by 300m, the recall of all approaches (even with ± = 0:0M) became 1.0. At the same time, as the size of an approximation be- comes larger, the number of false positives rises. When the sizeofthequeryrangebecomessmaller,theresultsapproach those of the point queries in Table 1. 6.3 Illustration of Directional Query Results: A Real-world Example We developed a web-based search system to demonstrate the feasibility and applicability of our concept of georefer- enced video search 1 . The search engine currently supports 1 http:==eiger:ddns:comp:nus:edu:sg=geo=Query idaho:html (a) Query points on map Label Query ID MBR Actual Vector (with di®erent overestimation constant) 0.0M 0.1M 0.2M 0.3M 0.4M 0.5M A 63 ° ° ° ° ° ° ° ° B 185 ° ° ° ° ° ° ° ° C 317 ° ° X X ° ° ° ° D 394 ° ° X X X X X ° E 465 ° X X X X X X ° F 740 ° ° X X X X ° ° G 761 ° ° ° ° ° ° ° ° °: found, X: not found (b) Query point overlaps using di®erent approaches Figure 12: Query points overlapped with an FOV (video 61, FOV 42). (a) (b) (c) (d) Figure 13: Impacts of bounding distance in video search. both directional and non-directional spatial range queries. A map-based query interface allows users to visually draw the query region and indicate the direction. The results of a query contains a list of the overlapping video segments. For each returned video segment, we display the corresponding FOVs on the map, and during video playback we highlight the FOV region whose timecode is closest to the current video frame on the map. InFigure14weillustrateanexampleofadirectionalquery applied to our real-world georeferenced video data. We would like to retrieve the video segments that overlap with the given rectangular region while the camera was pointing in the North direction. Figure 14.(a) shows the video seg- ments returned from the ¯lter step using the MBR model. Recall that the MBR model retains no notion of direction- ality. And Figure 14.(b) shows the results of the ¯lter step using the vector model with input direction 0 ± (i.e., North) and ± =0:3M. Figures 14.(b) and (c) show the results from the re¯nement step with error margins §5 ± and§25 ± with respect to the given direction 0 ± , respectively. Note that weappliedthere¯nementstepontheoutputofMBR-based ¯lter step shown in Figure 14.(a). Using the MBR model the ¯lter step returns videos with an aggregated duration of 775 seconds whereas the vector model based ¯lter step returns only 98 seconds of videos. There¯nementstepsshowninFigures14.(b)and(c)return 9 seconds and 65 seconds long videos for viewing directions 0 ± §5 ± and 0 ± §25 ± , respectively. TheFigures14.(a)through(d)alsodisplaytheFOVvisu- alizations for the corresponding video segments on the map. Note that a red FOV represents the frame currently being played in the video. As described in Section 3.1, the vector modelisintroducedasafast¯ltersteptoquicklydismissthe unrelated videos and video segments. Figure 14 illustrates an example query where the vector model successfully elim- inatesmostoftheunrelatedvideosegments, minimizingthe amount of re¯nement processing. 7. CONCLUSIONS Inthisstudyweproposedanovelvector-basedestimation modelofacameraviewablescenearea. Ourexperimentalre- sultsshowthatthevectormodelcanbeusedinvariousways toenhancethee®ectivenessofasearch¯lterstepsothatthe expensive and complex re¯nement step can be performed with far fewer potentially overlapping FOVs. The vector model successfully supports new geospatial video query fea- tures such as a directional point query with a given view- ing direction from camera to object and a point query with a bounded distance between camera and object. We also demonstrated an immediate applicability of our proposed model to a common database by collecting real video data, implementinganactualdatabaseandqueriesusingMySQL, and performing extensive experiments with the database. So far we have focused on understanding the feasibility of new query features in georeferenced video search using the proposed model. We did not investigate query optimization issues that can impact the performance of the ¯lter step, such as the most appropriate indexing structure. We are in- vestigating query optimization as part of our ongoing work. 8. REFERENCES [1] Flickr. http://www.°ickr.com. [2] self-reference removed for blind review processing. [3] Woophy. http://www.woophy.com. [4] N. Beckmann, H.-P. Kriegel, R. Schneider, and B. Seeger. The r*-tree: An e±cient and robust access method for points and rectangles. In ACM SIGMOD, 1990. [5] T. Brinkho®, H.-P. Kriegel, R. Schneider, and B. Seeger. Multi-step processing of spatial joins. In ACM SIGMOD, 1994. [6] Boris Epshtein, Eyal Ofek, Yonatan Wexler, and Pusheng Zhang. Hierarchical Photo Organization Using Geo-Relevance. In 15 th ACM Intl. Symposium (a) Results of ¯lter step using MBR model (no direction) (b) Results of ¯lter step using vector model (viewing direction=0 ± and ± =0:3M) (c) Results of re¯nement step (viewing direction 0 ± §5 ± ) (d) Results of re¯nement step (viewing direction 0 ± §25 ± ) Figure 14: Illustration of directional range query results. The FOV of the current frame is highlighted. on Advances in Geographic Information Systems (GIS), pages 1{7, 2007. [7] Shantanu Gautam, Gabi Sarkis, Edwin Tjandranegara, Evan Zelkowitz, Yung-Hsiang Lu, and Edward J. Delp. Multimedia for Mobile Environment: Image Enhanced Navigation. volume 6073, page 60730F. SPIE, 2006. [8] Clarence H. Graham, Neil R. Bartlett, John Lott Brown, Yun Hsia, Conrad C. Mueller, and Lorrin A. Riggs. Vision and Visual Perception. John Wiley & Sons, Inc., 1965. [9] Rieko Kadobayashi and Katsumi Tanaka. 3D Viewpoint-Based Photo Search and Information Browsing. In 28 th Intl. ACM SIGIR Conference on Research and Development in Information Retrieval, pages 621{622, 2005. [10] Lyndon S. Kennedy and Mor Naaman. Generating Diverse and Representative Image Search Results for Landmarks. In WWW '08: Proceeding of the 17th International Conference on the World Wide Web, pages 297{306, New York, NY, USA, 2008. ACM. [11] Xiaotao Liu, Mark Corner, and Prashant Shenoy. SEVA: Sensor-Enhanced Video Annotation. In 13 th ACM Intl. Conference on Multimedia, pages 618{627, 2005. [12] Mor Naaman, Yee Jiun Song, Andreas Paepcke, and Hector Garcia-Molina. Automatic Organization for Digital Photographs with Geographic Coordinates. In 4 th ACM/IEEE-CS Joint Conference on Digital Libraries, pages 53{62, 2004. [13] A. Orenstein. Spatial query processing in an object-oriented database system. In ACM SIGMOD, 1986. [14] A. Pigeau and M. Gelgon. Building and Tracking Hierarchical Geographical & Temporal Partitions for Image Collection Management on Mobile Devices. In 13 th ACM Intl. Conference on Multimedia, 2005. [15] Kerry Rodden and Kenneth R. Wood. How do People Manage their Digital Photographs? In SIGCHI Conference on Human Factors in Computing Systems, pages 409{416, 2003. [16] Ian Simon and Steven M. Seitz. Scene Segmentation Using the Wisdom of Crowds. In Proc. ECCV, 2008. [17] Carlo Torniai, Steve Battle, and Steve Cayzer. Sharing, Discovering and Browsing Geotagged Pictures on the Web. Springer, 2006. [18] Kentaro Toyama, Ron Logan, and Asta Roseway. Geographic Location Tags on Digital Images. In 11 th ACM Intl. Conference on Multimedia, pages 156{166, 2003.

Linked assets

Computer Science Technical Report Archive

Conceptually similar

PDF

USC Computer Science Technical Reports, no. 911 (2009)

PDF

USC Computer Science Technical Reports, no. 659 (1997)

PDF

USC Computer Science Technical Reports, no. 964 (2016)

PDF

USC Computer Science Technical Reports, no. 627 (1996)

PDF

USC Computer Science Technical Reports, no. 878 (2006)

PDF

USC Computer Science Technical Reports, no. 909 (2009)

PDF

USC Computer Science Technical Reports, no. 948 (2014)

PDF

USC Computer Science Technical Reports, no. 699 (1999)

PDF

USC Computer Science Technical Reports, no. 892 (2007)

PDF

USC Computer Science Technical Reports, no. 844 (2005)

PDF

USC Computer Science Technical Reports, no. 625 (1996)

PDF

USC Computer Science Technical Reports, no. 871 (2005)

PDF

USC Computer Science Technical Reports, no. 762 (2002)

PDF

USC Computer Science Technical Reports, no. 580 (1994)

PDF

USC Computer Science Technical Reports, no. 623 (1995)

PDF

USC Computer Science Technical Reports, no. 685 (1998)

PDF

USC Computer Science Technical Reports, no. 908 (2009)

PDF

USC Computer Science Technical Reports, no. 650 (1997)

PDF

USC Computer Science Technical Reports, no. 592 (1994)

PDF

USC Computer Science Technical Reports, no. 886 (2006)

Description

Seon Ho Kim, Sakire Arslan Ay, Byunggu Yu, Roger Zimmermann. "Vector model in support of versatile georeferenced video search." Computer Science Technical Reports (Los Angeles, California, USA: University of Southern California. Department of Computer Science) no. 912 (2009).

Asset Metadata

Creator Ay, Sakire Arslan (author), Kim, Sean Ho (author), Yu, Byunggu (author), Zimmermann, Roger (author)

Core Title USC Computer Science Technical Reports, no. 912 (2009)

Alternative Title Vector model in support of versatile georeferenced video search (title)

Publisher Department of Computer Science,USC Viterbi School of Engineering, University of Southern California, 3650 McClintock Avenue, Los Angeles, California, 90089, USA (publisher)

Tag OAI-PMH Harvest

Format 12 pages (extent), technical reports (aat)

Language English

Unique identifier UC16269472

Identifier 09-912 Vector Model in Support of Versatile Georeferenced Video Search (filename)

Legacy Identifier usc-cstr-09-912

Format 12 pages (extent),technical reports (aat)

Rights Department of Computer Science (University of Southern California) and the author(s).

Internet Media Type application/pdf

Source 20180426-rozan-cstechreports-shoaf (batch), Computer Science Technical Report Archive (collection), University of Southern California. Department of Computer Science. Technical Reports (series)

Access Conditions The author(s) retain rights to their work according to U.S. copyright law. Electronic access is being provided by the USC Libraries, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.

Repository Name USC Viterbi School of Engineering Department of Computer Science

Repository Location Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089

Repository Email csdept@usc.edu

Inherited Values

Title Computer Science Technical Report Archive

Description Archive of computer science technical reports published by the USC Department of Computer Science from 1991 - 2017.

Coverage Temporal 1991/2017

Repository Email csdept@usc.edu

Repository Name USC Viterbi School of Engineering Department of Computer Science

Repository Location Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089

Publisher Department of Computer Science,USC Viterbi School of Engineering, University of Southern California, 3650 McClintock Avenue, Los Angeles, California, 90089, USA (publisher)