Using RADIUS Site Models without the RCDE

Transcription

1 Reprinted (with minor corrections) from: Proceedings: DARPA Image Understanding Workshop, New Orleans (LA): Morgan Kaufmann Using RADIUS Site Models without the RCDE Aaron J. Heller, Christopher I. Connolly, and Yvan G. Leclerc Artificial Intelligence Center, SRI International 333 Ravenswood Ave., Menlo Park, CA USA Abstract We describe a system for converting RADIUS site models into a Web-accessible form. Once converted, the site models can be downloaded and viewed with standard, off-the-shelf Web browsers. Site models themselves are represented in the Virtual Reality Modeling Language (VRML). Each feature in the site model is cross-linked to a Web page that describes that feature in more detail, including image chips and collateral information. A demonstration of work described in this paper is accessible via the URL com/ connolly/pathfinder. 1 Introduction The explosive growth of the World Wide Web (WWW) in just three years has transformed it into a widely accessible medium for disseminating documents, sound, images, and recently, threedimensional models. This growth has also led to the development of a wide variety of tools for viewing and manipulating Web-accessible information. As a result, the Web is an attractive means of providing site model information. This paper describes a system for converting RA- DIUS site models into a Web-accessible form. After conversion, site models can be downloaded and This work was sponsored by SRI International. The views and conclusions contained in this document are those of the author and should not be interpreted as representing the official policies, either expressed or implied, of the Defense Advanced Research Projects Agency, the United States Government, or SRI International. viewed with standard, off-the-shelf Web browsers. Site models themselves are represented in the Virtual Reality Modeling Language (VRML). Each cultural feature in the site model is cross-linked to a Web page that describes that feature in more detail, including image chips and collateral information. 1.1 RADIUS Site Models As part of the RADIUS program, SRI has developed and assembled a suite of manual and semiautomatic tools for site-model construction that work within the RADIUS Common Development Environment (RCDE) [Heller and Quam, 1997, Heller et al., 1996]. Manual techniques are those in which the 3-D model of a feature is projected into one or more images and the operator adjusts the model to align it with was is seen in the images. Semi-automatic techniques are those in which an operator provides an initial rough estimate of a feature s position, size and topology and the system then refines or extends the model of the object using information extracted from the image(s). Figure 1 shows a portion of a typical RADIUS site model. The majority of the features modeled fall under three broad categories: Buildings and other structures such as bridges, and petroleum and water storage tanks. Lines of Communication such as roads, railroad tracks, and other linear features such as rivers and streams. Functional Areas such as parking lots, site Copyright c 1997 SRI International, 333 Ravenswood Ave., Menlo Park, CA USA All Rights Reserved.

2 perimeters, rail transfer points and other area features such as forested areas. Figure 2 shows a typical construction sequence for a building using the Model-Based Optimization system [Fua, 1996]. Similar techniques are available for extraction of linear and area features. 1.2 The Virtual Reality Modeling Language 1.3 What is VRML? The VRML Frequently Asked Questions (FAQ) 1 list says: VRML stands for Virtual Reality Modeling Language. It is usually pronounced V-R-M-L, but its friends pronounce it vermel. The goal of VRML is to create the infrastructure and conventions of cyberspace, a multi user space of many virtual worlds on the Net. VRML 1.0 is a subset of the Inventor File Format (ASCII) with some additions to allow linking out to the Web and including other URLs. The linking out feature (WWWAnchor) provides the same feature that HREF anchors provide in HTML. Another critical feature was the LOD (level of detail) which allows the right amount of data for an object based on how prominent it is in the scene, or the rendering speed of the browsing machine. Our current interest in VRML is a bit more pragmatic. While it is clearly not a replacement for richer scene description data models, such as the Synthetic Environment Data Representation and Interchange Specification (SEDRIS) [The SEDRIS Team, 1996] now being developed by the Defense Modeling and Simulation Office (DMSO), it is a simple and low-cost method to export geometric site-model data from the RCDE in a form that is suitable for dissemination via the WWW and usable with a wide variety of freely-available browsers and rendering tools History The development of VRML grew out of a discussion among several attendees of the first annual World Wide Web Conference in VRML was seen as a three-dimensional extension of HTML. As with most Web-related projects, the development of a VRML specification and prototype was rapid. A draft specification was released in fall of The format chosen for VRML was Silicon Graphics ASCII Open Inventor format. A specification for VRML 1.0 was finalized and released on May 26, 1995 [Pesce, 1995]. Not surprisingly, most of the initial work on VRML took place at Silicon Graphics. Over the next two to three years, between 20 and 30 VRML 1.0 browsers were developed for exchanging geometry over the Web Current status By August of 1996, a specification for VRML 2.0 was released [Bell et al., 1996]. While VRML 1.0 is capable of representing static geometry, VRML 2.0 was aimed at augmenting geometry with timevarying behavior. Because of its increased complexity, only 4 browsers are known to exist for VRML 2.0, as of this writing. The structure of a VRML scene description consists of an identifier string, to distinguish VRML 1.0 from VRML 2.0 descriptions, followed by a sequence of Node specifications. Each node represents a geometric feature or an attribute of that feature. For example, there are Cube, Sphere, Cylinder, and Cone nodes for representing those shapes. In addition, point sets can be specified. Faces (IndexedFaceSet nodes) are described as lists of points that form polygons. Material properties, certain transformations, viewpoints, and lighting properties can also be specified Strong Points An important advantage to using VRML 1.0 for disseminating site models is that many VRML 1.0 browsers have been developed across a wide variety of platforms. Although the VRML 1.0 specification has minor ambiguities, browsers exhibit generally uniform behavior, that is, VRML models will have the same or similar appearance across browsers. As 2

3 Figure 1: A portion of a typical RADIUS site model, projected onto an image of the site. The site model mainly consists of buildings, linear and area features. a result, VRML 1.0 has become a satisfactory, if imperfect, standard for exchanging geometry via the Web. VRML 1.0 also allows an association between objects and Web URLs. As with hypertext Web pages, mouse clicks on an object can be used to retrieve and display Web pages that are associated with the selected object Weak Points The main deficiency of VRML for our application is that there is no provision for georeferencing the data in the models. The VRML spec includes an Info node, which is intend to be used for comments. We press this into service to store the site-model s LVCS-to-geocentric-transform, to allow the recovery of geopositioning information and translation into standard cartographic coordinate systems. Unfortunately, VRML 2.0 removes some features that were found to be useful in VRML 1.0. Transformations have been restricted rather than generalized: 4 4-matrix transformations are permitted in VRML 1.0, but not in VRML 2.0. Moreover, in VRML 2.0, transformations cannot be sequentially specified and composed; they must be nested recursively, resulting in a bulkier representation for site models. As of this writing, there are no VRML 2.0 browsers that implement the complete specification. 3

4 This is in part due to the introduction of behaviors and scripting into VRML 2.0, which greatly complicates the semantics of the language. The VRML 2.0 standards process also appears to be dominated by commercial interests. As a result, VRML 2.0 is targeted mainly toward low-bandwidth home users browsing the Internet. The scientific visualization community appears to have had little influence in specifying the VRML 2.0 standard. 2 WWW Site Visualization Tool 2.1 The Translator The translator is implemented in approximately 3000 lines of Common Lisp that runs in the RCDE and operates on site-models that have been loaded into the system. Besides making the code somewhat simpler, this allows us to translate models regardless of whether they were loaded from RCDE featureset files, the RADIUS Testbed System database, or created in the current session. The translation process in implemented as a single pass over the objects in the site-model. Only those objects which are present in the selected view are translated. This allows the RCDE s feature set mechanism and associated menus to be used to select the set of objects to be translated. Every object in an RCDE site-model has an object-to-world-transform whichisusedtotransform the coordinates of the object s vertices to the site s local-vertical-coordinate-system (LVCS). This transform is translated into a VRML transform node. One complication is that VRML uses a righthand coordinate system in which the y-axis is up whereas the convention used by RCDE is that z-axis is up with the x-axis pointing east. In addition, any face of an RCDE object may have a texture map associated with it. The translator automatically creates the texture images and calculates the texture coordinates for inclusion in the VRML node describing the object. These texture maps can be either inlined in the VRML file or can be written into separate files that are referenced via a URL from the VRML file. 3-D objects that have a direct representation in VRML, such as cubes, cylinders, and spheres are simply translated into the corresponding node type. Other, more complicated objects, such as houses and complex buildings, are represented in the RCDE with a face-edge-vertex (f-e-v) datastructure, called the planar-solid class. This class of object is specialized by introducing parameterized constraints among the vertices and faces. The parameterized representation is used for adjusting the object and the f-e-v representation is use for drawing the object. Because of the availability of the f- e-v representation a single method suffices to translate all classes of compact 3-D objects into VRML indexed-face-sets. Roads and fences are represented by ribbon-curve objects in RCDE, which are comprised of a sequence of vertices that trace the centerline of the object and a width (or height for fences) at each vertex. VRML does not have a corresponding nodetype, so ribbons are triangulated and then translated into VRML indexed-face-sets. The RCDE represents terrain as regular quad-mesh or tri-mesh objects. Since the faces in a quad-mesh object are not necessarily planar, we use tri-mesh object for terrain and translate these to indexedface-sets. If a sun-direction vector is associated with the selected view, it is translated into a VRML DirectionalLight node and added to the scene file. 2.2 HTML Generation For each feature in the VRML representation, a URL is created containing an HTML page with attribute information and metadata for that feature. This page also contains image chips displaying the feature as it appears in all available images associated with the site. Image chips are generated by first collecting those site images within which the feature is visible. The extent of the feature in each image can then be determined by using the feature s bounding box, and transforming this (via the worldto-image transformation) into the corresponding image coordinates. The chip is created by windowing into the corresponding image, and converting this window into a GIF file for use in the HTML page. The feature attributes can be used to populate this page with appropriate text. The URL containing this page is then attached to the VRML representation by creating a WWWAnchor node linking the 4

5 feature to its own Web page. 3 Example Figures 3 through 5 show an example result. The initial page introduces the tool and lists the available site-models (Fig. 3a). The user has selected the Lockheed-Martin Corp., Denver site-model and is shown a page containing a synoptic view of the site and several alternative versions of the models (Fig. 3b). These have been generated to accommodate different network bandwidths and browser capabilities, ranging from small models that contain only the modeled objects, to fully texture-mapped models with high-resolution DTED. After selecting the version with texture and terrain, the VRML browser (in this case SGI s WebSpace) is started and the 3-D model is displayed (Fig. 4). At this point the user can fly around and inspect the model from any viewpoint. As the mouse is placed on the individual objects, they are highlighted and their names appear in a information window at the bottom of the browser. Clicking on any of the objects, causes the HTML browser to retrieve the page of attributes, metadata, and image chips for the selected object (Figs. 5 a&b). The Modifications and Comments fields on these pages, allow the user to enter data which is sent back to the HTTP server and ultimately incorporated into the site-model. 4 Related Work TerraVision II is the primary application of the recently started MAGIC II project. It is an extended version of the TerraVision terrain visualization application 2 that was created for the original MAGIC project 3. TerraVision was designed to visualize a single rectangular geographic area represented by elevation and orthorectified image data at multiple levels of detail. These data, typically larger than 1 Gb in size, are distributed across a high speed network. TerraVision accesses the data in realtime from the network as the user moves across the terrain, thus giving the illusion that all of the data is stored locally. This approach allows users access to very 2 magic/ terravision.html 3 large amounts of data without the need to copy, store, and maintain the data locally. TerraVision was inflexible because it was difficult to create a dataset of a very large area where different areas would have data at different resolutions. It was also impossible to use multiple types of images that could be fused together under user control. To overcome this inflexibility, TerraVision II is designed to use composite datasets consisting of many image pyramids, each of which can be created and stored independently at different sites. The glue that holds them together are VRML files. In a sense, TerraVision II will be an enhanced VRML browser that will be able to handle very large, network based, multi-resolution datasets. However, it will have additional capabilities for merging different image types (under user control) and it will produce seamless renderings of scenes with multiple levels of detail. We expect that an important source of data for TerraVision II will be RADIUS site-models produced by the VRML production system described here. 5 Conclusions Even though we have struggled with a number of deficiencies, ambiguities, and errors in the VRML specification and browsers, and are, in general, unhappy with the standards process and the direction in which VRML is evolving, we have still found it, and the freely-avaiable browsers, to be a useful mechanism for visualizing RADIUS site-models and illustrating possible methods for disseminating geospatial information on the WWW. Readers are encouraged to form their own opinions by experimenting with the example discussed in this paper. It is accessible via the URL sri.com/ connolly/pathfinder. References [Bell et al., 1996] Garvin Bell, Rikk Carrey, and Chris Marrin. The virtual reality modeling language specification version 2.0. URL: vag.vrml.org/vrml2.0/final/, [Fua, 1996] P. Fua. Cartographic Applications of Model-Based Optimization. In DARPA Image Understanding Workshop, Palm Springs, CA, 5

6 February Morgan Kaufmann. [Heller and Quam, 1997] Aaron J. Heller and Lynn H. Quam. The RADIUS Common Development Environment. In Oscar Firschein and Tom Strat, editors, RADIUS: Image Understanding for Imagery Intelligence. Morgan Kaufmann, San Mateo (CA), [Heller et al., 1996] A. J. Heller, P. Fua, C. Connolly, and J. Sargent. The Site-Model Construction Component of the RADIUS Testbed System. In DARPA Image Understanding Workshop, pages , [Pesce, 1995] Mark Pesce. VRML: Browsing and Building Cyperspace, chapter Appendix A: VRML: The Virtual Reality Modeling Language Version 1.0 Specification. New Riders, Indianapolis, Also available at the URL: [The SEDRIS Team, 1996] The SEDRIS Team. Synthetic environment data representation and interchange specification (sedris). URL: http: //

7 The original images. Sketch roof-line (Add Vertex). Done with roof-line (Drop). Correct elevation (MBO Z-Search). Optimize Shape (MBO Opt)... Done! Figure 2: The sequence of steps used to model a complex-shaped building with the extrusion primitive and the SRI-authored Model-Based Optimization system. This entire sequence typically takes less than one minute of elapsed time. 7

8 (a) (b) Figure 3: The homepage for the WWW Site Visualization Tool and initial page for the Lockheed-Martin Denver site. The URL is connolly/pathfinder Figure 4: A texture-mapped, 3-D rendering of the site-model created from the VRML description of the site and displayed by SGI s WebSpace. 8

9 (a) (b) Figure 5: The a sample of the Web pages that are hyperlinked to every cultural feature in the site model. The text and image chips are generated automatically from the attributes, metadata, and images stored with the site-model in the RCDE. 9