Making Image Mosaics Infinitely Zoomable to Explore Large Image Collections

Transcription

1 Making Image Mosaics Infinitely Zoomable to Explore Large Image Collections Romain Vuillemot Béatrice Rumpler Université de Lyon LIRIS F-69621, France ABSTRACT Image Mosaics are images formed from series of smaller images. They are inspiring, intuitive and trigger curiosity. But remain still images. In this paper we introduce a pure content-based interaction model to make them interactive and infinitely zoomable. Our goal is to provide users a smooth and multi-scale information organization to better visually explore large non-structured image collections. Experiments on a non-familiar dataset indicates it is a smooth an efficient narrative environment. Guidelines and implementation notes are given to extend to any mosaic interface, such as non-squared mosaics and collages. ACM Classification Keywords H5.2 Information interfaces and presentation: User Interfaces. - Graphical user interfaces. General Terms Image Mosaic, Vizod, ZUI Author Keywords Image Mosaic, Zoomable User Interface, Visualization on-demand INTRODUCTION Informal observations of people walking down the street showed that when passing by an image mosaic, people tend to stop and pay attention to it. Catching this attention span is priceless in our information society, where relentless streams of information are unwillingly dumped into our brain, through our eyes. Image mosaics are part pieces of art, part technique, and have traveled through the ages, and regained of interest during last decade with the rise of the digital era. Even if this spatial organization of images is sophisticated, everybody knows how to get information out of it, regardless age, culture or education. Source images can be fetched in a 2D way as tiles, clustered by color or shape, according to the big or target image that is visible by stepping backward. Meanwhile, it has never been so easy to capture images and make them available to a global community of people. Images tend to gain better quality and higher size since displays and captors follow a restless trend in being better and cheaper. But studies [18, 20] show that people don t invest much time classifying their photos, relying on automatic metadata such as photo transfer time or cameras timestamps. Internet and social websites now offer open repositories of images, that are a goldmine for casual image lovers, but a frustrating experience for people in need of a particular information. Even with an efficient information retrieval system, browsing large amount of results is not there yet. For instance, seeking for a landscape picture, users will start by a broad term and then rely on browsing to find the right image. But a query using that keyword is resulting in Flickr by: 2,341,040 results matching landscape. How is it possible for a human being to explore such a large amount of data in a limited time span? Because screens size remain limited, a better use of human visual bandwidth has to be found. While users are eager of image tiles [18, 6], our conviction is one step beyond, that image overload is a key to better use human visual ability, and can be harnessed by a strong control of the interface. Filling screen space with images in a meaningful and understandable way, is the underlying challenge of any image manipulation interface. Our focus is on the worst case scenario where images are not familiar and have absolutely no information that may index them, but by content. Spatial arrangement has been experimentally proven efficient [17], so are humans spatial memory abilities [5]. But when it comes to deal with photos becoming too small to recognize, reductions by representative photo [15] or density plot [13], the system may miss image features. Collage are nice image display [10] but cluttering appears and remains a major issue. And cluttering is not only about showing images with no overlapping, but also not to lead to a degradation of performance at some task [19]. In this paper our hypothesis is that image mosaics are an intuitive and efficient spatial organization of images, that can pack a vast amount of images with minimal degradation. Images are indexed based on a matching between their own content (shape, feature, color,..)

2 Figure 1. We introduce MosaiZ, infinitely zoomable image mosaics. Starting from (a), the user pans mosaics to look for a target image that is a half dark, half clear. But no satisfactory image is found, so the user decides to make a visual query by zooming on a part of the current image that his fulfilling his needs. Border of the bridge with the sky seems to be a good candidate query, so the user is zooming on it. The image is becoming a mosaic itself (b). And zooming again (c) is showing more results, and result images resolution gain in better quality (d) (user is hovering it, mouse is not drawn in the screen capture). and the target image content: they can easily be visually acquired by looking at the target image. Here are characteristics of image mosaics: Space filling organization with an optimal use of screen space. Every pixel holds informations on both source and target image. No clutter or image overlapping. Even with many images, they remain strictly adjacent to each other. Both overview and details at a glance. Image mosaic can be considered as a non-photo realistic rendering process of the target image by the source images. Reificating real world metaphor is a common and advised practice in UI design [4]. While Informative Art [16] is using artistic visualization to dynamically display updated information, image mosaics can be an advanced content browsing system. Even if through the years mosaics have been studied and improved using various angles (shape optimization, algorithm optimization,..), as far as we know they have never been studied as a browsing metaphor. And since the digital world is not limited by space and distance, it is possible to go one step beyond reality and make mosaics infinitely zoomable. Using a Zoomable User Interface (ZUI) is straightforward and has already been successfully used with image management [2]. Zooming into mosaics is based on the property that images are made of images themselves, becoming an infinite two dimension plane. Semantic zoom [14] is automatic : when the source image miniature becomes too small, the target image appears. Here are how we are tackling major design challenges: Consistent and predictable layout due the nestedness property of the zoomable mosaics. The transitional step (zoom), from source to target (or the opposite) is a global linear geometric scaling -and no hyperbolic deformation such as Fisheye. Magnification is the same for every images. Humans know well to deal with and anticipate such deformation. Scalability. A target image is summarizing the source dataset. In the scenario where source images are one pixel by one pixel, nearly 2 million images can be packed in a pixel image. Also mosaics can be easily exploded into manageable chunks of images, allowing progressive streaming over a network. Users know how the system works. Because image mosaics are based on visual perceptions properties, there are no cultural or context based knowledge (such as classification or clustering). MOSAIZ Our contribution is called MosaiZ, a pure content-based interaction model to make current image mosaics interactive and infinitely zoomable (See Figure 1). A MosaiZ is a non-linear interaction space, where users can select a source image that will become the target image, and so on. Maximizing information display is the priority, to display as much information as possible. The design has to be pure content-based with no menu or buttons. Interactions aim at helping users to visually formulate his query (image he is looking for) using target image features. This way users express their needs in an understandable way for the machine, with no sign or code. Smooth transitions (continuous zooming) helps keeping context. We describe guidelines to create a MosaiZ and then the interaction design. Creating a MosaiZ The backbone of a MosaiZ is a scene graph (similar as a space scale diagram) connecting image mosaics altogether. Every nodes of the graph is a mosaic, and will be connected to all the sources that are composing that mosaic. Here are the steps to follow: 1. Select a dataset and a first image as scene graph root 2. Select one grid type, and one or many sizes

3 3. For each grid size: (a) Generate mosaic (b) Assign original image to source images 4. Check connectivity of the MosaiZ graph Interaction Design We combined interaction techniques in a meaningful and consistent way to help users to make the best decision at every level in the graph. Animation helps to better understand relationships by maintaining context awareness in a single view using short term memory: there are no attention split or gap between views. The user interaction workflow is described on Figure 1. Here are some specific design details regarding mouse and keyboard: Mouse We implemented a classic panning (translation) and zooming (uniform scale changes) navigation. Scroll up is to zoom in a location, and scroll down to go backward in the multi-scale space. When the mouse exits the application window, original image is displayed to ease the feature seeking process (Figure 1 (a)). When the mouse enters, the mosaic is on but with a 50% alpha level couple with the original image (this is a common trick used in image mosaics) that helps to better perform visual query (Figure 1 (b)). This cheat feature (so called in the metapixel library we will be using in the implementation) may be disturbing and biais user decision. So when mouse is hovering the target image, the original source image is displayed (Figure 1 (d) but without vicinity) and its vicinity as well in a transitional way (Figure 2). Figure 2. Mouse pointer is an extra layer that helps users to get better source images resolution. Multiple shapes are available with variations in size and transparency level (the darker the square is, the better the quality of image is). Keyboard Arrow keys are to pan around, but zooming is difficult because target acquisition is not easy since the system has no information about users visual focus. Our solution is by using the numeric pad, users progressively select areas of the screen (Figure 3). We introduce in the next section automatic touring that is an automatic straight zooming, and that can change direction according to the key pressed, such as a to pilot a plane: it reduces users interaction, but requires users attention. Figure 3. Image target acquisition through the keyboard is performed by pressing numeric pad keys (7, 9, 1 and 3). Keys are tied to one quarter of the screen. Every time a key is pressed, the region available is a quarter of the previous one. Features We now describe four major features to complete basic interactions we described above. Other features are given in the users experimentation section. History Any system must give users the opportunity to make mistakes and go backward. Users can right click to go to the previous state (previous scene graph node). But also visual clues such as red squared images or making target image reddish shows already visited ones. History or statistics can also be exported such as list or graph to be browsed in another interactive environment. Mosaic Grid The grid is the mosaic pattern making the basic support to construct mosaics. Dynamically changing the mosaic grid size is a powerful feature since it will allow to change the query bounds. Grid resizing is done by using mouse scroll and pressing control key at the same time. Grid sizes have to be set during the image mosaics generation step. Rendez-Vous Point Every exploratory tasks involve non predictable, confuse and irrational behaviors. Sometimes users just want to find by serendipity or to have fun with the tool, rather than following beaten tracks. For that purpose, we allow users to set a Rendez-Vous Point : users press a key at some point, explore, and by releasing the key will go back to the Rendez-Vous. No history is recorded. The way back to the Rendez-Vous is a fast automatic animation to quickly remember what has been done during that recreation time. Letting users pressing the key is important to remind them that this is just a short term image exploration process. I am Feeling Lucky Systems -and especially new ones- should not let users stuck in a situation where they can t do anything or don t know what the next step is. In a MosaiZ, spatial freedom and zoomable process require lots of user visualization and concentration. Situations such as desert fog [9] -being lost in the multi-scale space with not enough information to take a decision- have to be tackled. We introduced a system-based strategy to prevent it. By double clicking (instead of single click),

4 the system will look for the next image, regarding a user-selected heuristic that may be browsing new images only, performing the same query as the previous but on a new image, browsing by chronological order, image size, social popularity.. This automatic touring feature will free users from interactions. It can also be seen as a shortcut to repetitive actions users perform. We also introduce a semi-automatic approach, automatic straightforward zooming, users pressing arrow keys or moving mouse to set the direction. The result is a time dependent interface such as Dasher [23] which is using letters flow to compose phrases. In that case user s attention is required and is parametrable: t touring = s (t t start ) with time-lapse visualization (s < 1), real-time visualization (s = 1) or slow motion visualization (s > 1). Figure 5. Control panel allows users to customize interactions with the MosaiZ. IMPLEMENTATION NOTES We implemented MosaiZ based on interaction design and features we described in a twofold approach (Figure 4): 1. A local Java application using Picollo2D (open source project based on [3]) which is a direct-manipulation graphics library that supports constructing ZUI. The application is a basic frame containing a layer on which three child nodes (mosaics) are tied (previous, current and next mosaic). Pointer layer is added and tied to mouse cursor, to display child images according to the pointer location, size and shape. A set of event listeners trigger node rotation when users get close enough to an image (scale > 1) or click on it. We used mipmaps (pre-calculated, optimized collections of images) as a cache to quickly provide images into the application, and to prevent aliasing effects. MosaiZ implementation code contains a dozen of classes and a total of 1200 lines of code and 2 shell scripts. A modular implementation makes it easy to add touring new heuristics or other mosaic grids. 2. A remote server is generating the scene graph, based on image descriptors uploaded the by client application. Server is keeping track of users activity, based on [22] and is reached as a mosaic generation web service. Since image mosaic generation is not fully predictable, we wanted to avoid bottlenecks in user experience, and generated the graph scene once for all in a offline mode. We used Metapixel to generate the mosaics with different grid sizes. Activity logs (indicators such as path in the MosaiZ, frame rate, session times, features usage..) are uploaded on the server for analysis after every sessions, in order not to bias the performances. Regarding the technical challenge that implementing MosaiZ is, we started using a relatively small dataset to circumnavigate technical limits (memory and mosaics generation). We are using as dataset a selection of 250 photos from New York City with heterogeneous subjects (buildings, animal, people, textures, indoor, outdoor, flash, no flash, landscape,..). Pictures are homogeneous in resolution and compression, taken by the very same person (which is none of the testers) with a SLR camera. Pictures were rotated if needed, resized and cropped to 400 pixel by 400 pixels. Equipment A commonly available computer system was used, 3Ghz Intel PCs with 2Go RAM and a 21 monitor, running Windows XP. Screen resolution was of Interactions were performed using a 3-button mouse with scroll wheel, and a basic keyboard. All users and results mentioned below are based on this environment. Experiment #1 with a Digital Artist A professional male digital artist, with a 3-year experience background in design of websites and brochures, has been contacted and volunteered to give informal feedbacks using the prototype. A time expensive and boring part of his daily routing is finding adequate pictures from professional image repositories. Image are high quality and copyright issues are a major concern. Seeking focus is not on a global scene, rather on parts of the image, or texture that will be used in a montage or as a backgrounds. We briefly introduced him the interface and he instinctively started using it, for a 10 minutes session. Having a teaching and collaborative work experience it was easy for him to tell his thoughts at the same time while being focus on the application. EXPERIMENTATION The research question we want to answer is how MosaiZ are useful in the context of image collection browsing? We performed a qualitative user study with both a digital artist and knowledgeable users. Results His major satisfaction was not starting from scratch and having a basic ground of exploration, by starting with

5 Figure 4. Twofold MosaiZ implementation architecture. Local interactive client remains on users side (left). A connexion to a remote server (right) allows (strates from top to bottom): session opening, features upload, scene graph download and finally users interactions. Logs are uploaded when session ended, and users access to their browsing history using an ordinary web browser. one image. Not having that blank sheet fear, was a creative first step to him. He suggested taking himself a picture or drawing as a good start. As a game lover, he enjoyed the gameplay and extensively used scrolling and panning. Features such as history and feeling lucky haven t been used at all, because he was mostly relying on his good visual and spatial memory he developed over his experience. He liked the rendez-vous, which is a common feature in Photoshop where drawing layers can be removed or aggregated, allowing history visualization and transformation. The only customization he made was replacing clicking from target acquisition, to progressive zooming, which gave him more control over the interface. He thought it was cool but needed better quality, larger and more familiar dataset to really compare to his needs. But he would use it as a presentation medium because of its innovative design and continuous process. Experiment #2 with Knowledgeable Users Four Computer Science PhD candidate volunteer took part in a second experiments. They have responded to requirements of using image manipulation software at least once a week and with large image collections. Participant were asked if they were familiar with the dataset, one claimed he knew the city. Some statistics have also been recorded and are discussed in the discussion section. An interface demonstration has been given to the participants, they were observed during the execution, and asked to think loudly during a 10 minutes session. Results The first impression of all participants was that it was cool and always wanted to do it. The participant who claimed to be familiar with NYC was the one who developed the most observation, telling himself stories based on memories. His exploration process was based on the semantic of the image. While other participants rushed to try the usability of the interface, and focused on low level features (shapes, color,..). One participant particularly interested in skyscrapers found the interface very useful: by clicking on horizons or building borders with the sky, another building pictures appeared every time. Another candidate used the Rendez-Vous point as a quick magnification tool by being quickly activating and deactivating it. This was also used as a zooming out, which is similar to the Exposé feature in MacOS X. Participants found the export feature useful because it helped them to remember and get details on the image they were looking for. They wanted to annotate and record Rendez-Vous Point, and a replay feature that

6 would display text or voice annotation based on their MosaiZ path. But even with history visualization, participants complained that it was hard to be sure to complete their own personal challenge of complete browsing. All participants extensively used the feeling lucky button because they did not want to make decision and also because it was hard to visually explore and make up their mind. Participants comments were mainly about the limit of query by existing content. But still were satisfied to have a continuous flow of information and not waiting, always having the feeling to be in control. A participant compared the interface as the ipod where the wheel makes the navigation enjoyable. participant session. Not all the images seem to have been browsed. One major reason is the distribution of the images in the mosaics: some images are good candidates and appear multiple times in mosaics. We checked the connectivity of the scene graph and some graph analysis may also explain participants behavior. A solution would be to constraint the mosaic library, but this is a trade-offs between time, aestheticity and usefulness. Connectivity is also better difficult with smaller grids, but it already hold most of the pre-computation resources (Figure 7). Participants used the control panel (Figure 5) to customize the interface. Displaying grid was one of the favorite feature, that is a one-pixel black border around images for delineation. Indeed images mapping blue sky are not easy to separate. Changing grid was a deception since bigger sizes did not make recognizable mosaics. Still with the pixel grid, participants liked having similar images in the same vicinity. Some interesting behavior appeared, such a participant wanting to close and restart a session with no reason. Even if the space is infinite, some closure process or fresh started were needed. DISCUSSIONS In our two experiments, all five participants definitely enjoyed the experience and the learning time was very short. Participants reaction was also based on their own sensibility. Apart the digital artist, participants felt comfortable by reducing there degrees of freedom, being sure they were not missing something. Participants seems to be better at examining lists rather than 2D displays, but MosaiZ offer contextual awareness that provides more information overload. Figure images mosaics are generated according to 6 different grid sizes. The smallest one, the 10-pixel size grid mosaics are the one requiring most the the computation time. Performances were a major concern in our approach. While interacting with a visual display, the frame rate is a good indicator about the smoothness of an application. We noticed below 15 frames per second a movement doesn t seem to be smooth anymore. Results (Figure 8) show some drops in the rate that are tied to a user zooming action to a new mosaic. There is a design enhancement to make in the transitional state (See Figure 1) from (d) to (a). In (d) there are only two images being rendered whereas in (a) many other (surrounding the target) and in bigger size and quality. But participant never complained about it. Offline performances concern image mosaic generation: it took 2h32mins to generate all the mosaics. Globally with know extra device we learned on users eye gaze activity and feedback. The task to perform was limited, dataset was not significant but users appreciated the browser-like design of the MosaiZ interaction model. Figure 6. Peripheral target visualized in a 10 minutes session. 2-peripheral image means we consider images 2 grid units away from focus. In that case, since content is of 250 images, 220 images (88%) have been browsed. On Figure 6 we display the image browsed during one FUTURE WORK Our next iteration is to make the program technically scalable up to million of images. The bottleneck will be mosaic generation, but can be outsourced and computed in parallel on a grid or a cloud, because they are independent processes. Next research tracks are mainly bridging links with other communities such as image analysis, computer graphics and social networks:

7 software are available and are free such as KMosaic issued from [12] and MetaPixel are fully usable. Collage are mosaics with pieces having different shapes, that can be rotated, overlapping, such as we took a scissor and cut magazine to make a new pictures. Finally QuadTree based on image segmentation are well known and implemented technique that can be integrated as a grid. And pointer shapes can segment areas, instead of remaining constant shape, that will assist users providing automatic target similarities detection. Figure 8. Frame rate during a 60 second session sample. Every drop to 11s, 12s, indicate zooming interaction. 100 is the maximal frame rate and is reach for all the other activities. Context According to participants, MosaiZ would to be good when dealing with outputs from a previous system, such as the results of a Information Retrieval system. The major limit is the very long offline mode that is required and would not make MosaiZ a real time exploratory tool. It will be efficient with large results dataset but would not work below a number of results. Indeed a minimal quantity and quality of images are required to make good mosaics. Collaboration and social experience A MosaiZ instance (i.e. path in a MosaiZ) can be shared experience, and used as photo presentation. Through history it will be possible to adjust path, by un checking unwilled photos. Using other modalities is an interesting path of study, such as vocal annotation, music. Experiences gathered through massive MosaiZ usage may reinforce the feeling lucky heuristic. This social feedback may also be integrated into social image mosaics (e.g. a color code is showing the most frequently clicked targets and result in heat map). Above storing and retrieving tasks, communication and presentations are also important features when it comes to deal with image collections. Storytelling [1] is a feature expected by image authors. As recommended by our digital artist, a presentation tool can be made out of it, with contentbased transitions. Using a zooming approach has been already been successful with CounterPoint [8]. Other grids and pointing shapes The grid changing feature is interesting, but so many mosaics drawings and image segmentation algorithms are available and can help to do better than squares. Plenty of contribution are available, from SIGGRAPH papers to student projects. Numerous mosaic types have already been object of study [7, 21] during last decade. Anti-Mosaics are classic mosaics but images are the parts of a single image. Recently introduced Jigsaw Image Mosaic [11] are based on a dataset that has shape similarities and a shape homogeneity. Some CONCLUSION We described MosaiZ, an innovative interface to explore image collection by making image mosaics interactive and infinitely zoomable. Even if packing a lot of images into a delimited area is a challenge, we provided a scalable and smooth interface in which can be done visual queries. A wide range of optimizations and trade off are available, due to the wide design space. Our implementation and experiments were limited but comforted us in our initial hypothesis that image mosaics are a promising use of human visual bandwidth. Further work has to be done to make it fully usable and solving real tasks. Another exciting challenge is to federate communities to work together to extend the generic interaction model to other mosaic creation and image segmentation techniques. ACKNOWLEDGMENTS We thank all the testers, and developers of the Piccolo and metapixel libraries. REFERENCES 1. Marko Balabanović, Lonny L. Chu, and Gregory J. Wolff. Storytelling with digital photographs. In CHI 00: Proceedings of the SIGCHI conference on Human factors in computing systems, pages , New York, NY, USA, ACM. 2. Benjamin B. Bederson. Photomesa: a zoomable image browser using quantum treemaps and bubblemaps. In UIST 01: Proceedings of the 14th annual ACM symposium on User interface software and technology, pages 71 80, New York, NY, USA, ACM. 3. Benjamin B. Bederson, Jesse Grosjean, and Jon Meyer. Toolkit design for interactive structured graphics. IEEE Trans. Softw. Eng., 30(8): , Alan F. Blackwell. The reification of metaphor as a design tool. ACM Trans. Comput.-Hum. Interact., 13(4): , Andy Cockburn, Carl Gutwin, and Jason Alexander. Faster document navigation with space-filling thumbnails. In CHI 06: Proceedings of the SIGCHI conference on Human Factors in computing systems, pages 1 10, New York, NY, USA, ACM.

8 6. Tammara T. A. Combs and Benjamin B. Bederson. Does zooming improve image browsing? In DL 99: Proceedings of the fourth ACM conference on Digital libraries, pages , New York, NY, USA, ACM. 7. Adam Finkelstein and Marisa Range. Image mosaics. Lecture Notes in Computer Science, 1375:11??, Lance Good and Benjamin B. Bederson. Zoomable user interfaces as a medium for slide show presentations. Information Visualization, 1(1):35 49, Susanne Jul and George W. Furnas. Critical zones in desert fog: aids to multiscale navigation. In UIST 98: Proceedings of the 11th annual ACM symposium on User interface software and technology, pages , New York, NY, USA, ACM. 10. Andruid Kerne. Collage machine: Interest-driven browsing through streaming collage. In In Proceedings of Cast01: Living in Mixed Realities, pages , Junhwan Kim and Fabio Pellacini. Jigsaw image mosaics. ACM Trans. Graph., 21(3): , Kihwan Kim, Irfan Essa, and Gregory D. Abowd. Interactive mosaic generation for video navigation. In MULTIMEDIA 06: Proceedings of the 14th annual ACM international conference on Multimedia, pages , New York, NY, USA, ACM. 13. Zoran Pecenovic, Minh Do, Martin Vetterli, and Pearl Pu. Integrated browsing and searching of large image collections. In VISUAL 00: Proceedings of the 4th International Conference on Advances in Visual Information Systems, pages , London, UK, Springer-Verlag. 17. Kerry Rodden, Wojciech Basalaj, David Sinclair, and Kenneth Wood. Does organisation by similarity assist image browsing? In CHI 01: Proceedings of the SIGCHI conference on Human factors in computing systems, pages , New York, NY, USA, ACM. 18. Kerry Rodden and Kenneth R. Wood. How do people manage their digital photographs? In CHI 03: Proceedings of the SIGCHI conference on Human factors in computing systems, pages , New York, NY, USA, ACM. 19. Ruth Rosenholtz, Yuanzhen Li, Jonathan Mansfield, and Zhenlan Jin. Feature congestion: a measure of display clutter. In CHI 05: Proceedings of the SIGCHI conference on Human factors in computing systems, pages , New York, NY, USA, ACM. 20. Ben Shneiderman, Benjamin B. Bederson, and Steven M. Drucker. Find that photo!: interface strategies to annotate, browse, and share. Commun. ACM, 49(4):69 71, Nicholas Tran. Generating photomosaics: an empirical study. In SAC 99: Proceedings of the 1999 ACM symposium on Applied computing, pages , New York, NY, USA, ACM. 22. Romain Vuillemot, Béatrice Rumpler, and Jean-Marie Pinon. Anatomy of a visualization on-demand server - a service oriented architecture to visually explore large data collections. In ICEIS (5), pages 86 93, David J. Ward, Alan F. Blackwell, and David J. C. MacKay. Dasher - a data entry interface using continuous gestures and language models. In UIST, pages , Ken Perlin and David Fox. Pad: an alternative approach to the computer interface. In SIGGRAPH 93: Proceedings of the 20th annual conference on Computer graphics and interactive techniques, pages 57 64, New York, NY, USA, ACM. 15. John C. Platt, Mary Czerwinski, and Brent A. Field. Phototoc: Automatic clustering for browsing personal photographs. Technical report, Johan Redström, Tobias Skog, and Lars Hallnäs. Informative art: using amplified artworks as information displays. In DARE 00: Proceedings of DARE 2000 on Designing augmented reality environments, pages , New York, NY, USA, ACM.