Proc. 140th SMPTE Technical Conference, 1998
`Adding Hyperlinks to Digital Television
`V. Michael Bove, Jr., Jonathan Dakss, Stefan Agamanolis, Edmond Chalom
`MIT Media Laboratory
`Cambridge, MA USA
`Abstract
`Hyperlinked video is video in which specific objects are made selectable by some form of user
`interface, and the user’s interactions with these objects modify the presentation of the video.
`Identifying and tracking the objects remains one of the chief difficulties in authoring hyperlinked
`video; we solve this problem through the use of a video tracking and segmentation algorithm that
`uses color, texture, motion, and position parameters. An author uses a computer mouse to
`scribble roughly on each desired object in a frame of video, and the system generates a
`segmentation mask for that frame and following frames. We have applied this technique in the
`production of a soap opera program, with the result that the user can inquire about purchasing
`clothing and furnishings used in the show. We will discuss this and other uses of the technology,
`describe our experiences in using the segmentation algorithm for hyperlinked video purposes in
`both broadcast and on-demand modes, and present several different user-interface methods
`appropriate for hyperlinked video.
`
`Hulu
`Exhibit 1011
`Page 0001
`
`

`

`Introduction
`Users of the World Wide Web have become accustomed to hyperlinks, in which “clicking” on a
`word or graphic selects another page, or perhaps modifies the current one. The idea of
`hyperlinked video, in which objects are selectable, has often been discussed as a desirable
`possibility – consider for instance a fashion program in which clicking on an article of apparel
`provides information about it, or a nature program that lets children go on “safari” and collect
`specimens. Playback of such material is well within the capabilities of typical digital television
`decoders with graphical overlay capability, but creating it has posed a challenge because of the
`tediousness of identifying the clickable regions in every frame manually and the difficulty of
`segmenting and tracking them automatically.
`We have developed a method for tracking and segmenting objects which simplifies the authoring
`problem. During editing, the author simply uses the mouse to scribble roughly on each desired
`object in a frame of video, and the system generates a filled-in “segmentation mask” for that
`frame and for following and preceding frames, until there is a scene change or the entrance of
`new objects. Regions can then have an action (e.g. graphical overlay, switching to a different
`video data stream, transmission of data on a back channel) associated with them. The viewer can
`select objects by a variety of means, including ordinary remote controls or remote controls with
`inertial sensors; in our demonstrations we have set up a video projector that can identify and
`locate a laser pointer aimed at its screen.
`Our hyperlinked video system has two main technical underpinnings. First, we apply a novel
`method of using color, texture, motion, and position for object segmentation and tracking of
`video. While the author is indicating regions of importance by scribbling with the mouse, our
`system uses a combination of these features to develop multi-modal statistical models for each
`region. The system then creates a segmentation mask by finding entire regions that are
`statistically similar and tracking them throughout a video scene. The second piece is a scripting
`language for object-based media called Isis, which enables the association of the mask with the
`video frames, and the association of actions with selecting regions at particular times.
`We utilized this system to author a prototype hyperlinked video program which resembles an
`episode of a television serial drama and enables the viewer to select props, clothing and scenery
`and be shown purchasing information for the item, including the item’s price and retailer. We
`produced the content used in the program, which we named “HyperSoap,” with the belief that
`existing television material could not be adapted effectively to hyperlinked video. This entailed
`changes in the way HyperSoap was scripted, shot, and edited when compared to a traditional
`television program. We were able to learn a great deal about how people interact with
`hyperlinked video, and based our design of several modes of user interaction on this information.
`In the next section we will describe existing hyperlinked video software. In following sections
`we will describe the issues we were confronted with in designing systems for authoring and
`playing back hyperlinked video. We then will discuss the design of video content for a
`hyperlinked video production, using HyperSoap as an example. Finally we will summarize our
`work and present other applications for this technology.
`Authoring Hyperlinked Video
`Several companies have announced products for hyperlinked video authoring. VisualSHOCK
`MOVIE from the New Business Development Group of Mitsubishi Electric America, Inc.
`(http://www.visualshock.com) concentrates on Web and local playback (e.g. CD-ROM, DVD-
`ROM). The authoring tool, called the Movie MapEditor, requires specifying a rectangular
`
`Hulu
`Exhibit 1011
`Page 0002
`
`

`

`bounding box around the location of a desired object in the starting and ending frames of a video
`sequence, and a tracking algorithm estimates the position of the “hot button” in intermediate
`frames; manual tracking by the operator is also supported. HotVideo from IBM’s Internet Media
`Group (http://www.software.ibm.com/net.media/solutions/hotvideo) has a similar approach, but
`supports a larger set of geometrical object shapes. Intermediate locations of an object may be
`specified by multiple keyframes or located by a tracking algorithm. Another current authoring
`tool to incorporate an object-tracking algorithm is Veon’s V-Active (http://www.veon.com).
`We have found that this method for associating links or actions to objects is too coarse for the
`objects found in many varieties of video sequences. To enable the association of links with
`arbitrarily shaped regions, a system is needed which can differentiate among objects with higher
`precision.
`Segmenting Objects
`If we want to enable the author to specify clickable regions that follow object outlines rather than
`simple geometrical shapes, and if we don’t want to require manual outlining in every frame,
`some automated segmentation method is necessary. For example, the VisualSEEk system [1]
`automatically identifies regions based on color and texture distribution. This and other automatic
`systems are limited by the ability of the system to generate an internal model that corresponds to
`actual objects as required by the author. Good quality is important for situations in which the
`labels the system generates for pixels (known as a “segmentation mask”) are rendered for the
`viewer, for example to indicate the clickable regions. It must be noted, of course, that playing
`back video with a detailed segmentation will require much more information to be transmitted or
`at least retained on the server than would be the case with geometrical regions that can be
`described in terms of their coordinates.
`We have developed a novel segmentation system which classifies every pixel in every frame of a
`video sequence as a member of an object, or group of objects, as defined by an author [2][3].
`The system enables the author to define objects by labeling representative clusters of pixels in a
`single frame. The pixels which the author has classified are then tracked through the remaining
`frames of the sequence. The system computes multiple features for all of the pixels in the
`sequence, which include color, texture, motion and position. For each frame, statistical models
`for each object are created based on the features of the classified pixels. These models are then
`used to classify each unlabelled pixel to the object with highest statistical similarity.
`We have created a simple drawing tool which enables the author to indicate regions of
`importance within a video sequence. The user selects a single representative frame from the
`sequence (typically one in which the important regions are fully visible) and uses the mouse to
`highlight representative pixels for each desired object. When the author has finished labeling
`pixels in the frame, our system estimates the location of these pixels within each of the
`remaining frames in the sequence using block matching. When this stage has completed, there
`are pixels in each frame which have been classified to correspond to objects. The next stage of
`processing will assign the rest of the pixels in the image to one of these objects.
`Starting from small subset of pixels that the author has indicated in one frame and classifying
`every pixel in every frame is done by calculating a multi-modal statistical model based on a
`variety of different features including the color, motion, and position of a pixel and the texture
`and motion properties of a small neighborhood centered upon it. We have found that this yields
`a more robust segmentation, whose quality is minimally affected by misleading data from a
`single feature. Because a region’s model is multi-modal, it is for instance possible to specify a
`
`Hulu
`Exhibit 1011
`Page 0003
`
`

`

`region that has red and yellow pixels, and the system will not assign intermediate shades of
`orange to that object.
`After we have calculated a multi-dimensional feature vector for every pixel in the video
`sequence, we use the feature information for the tracked pixels to form a statistical model for
`each user-identified region. We then compare the feature information of each unclassified pixel
`to these models and label the pixel as a member of the region to which it bears the greatest
`statistical similarity. Basically, we use feature information to express each region as a
`mathematical function such that we can input feature values of unclassified pixels and receive as
`output the probability that the pixel corresponds to that region.
`Because only statistical methods are employed to classify pixels, it is likely that there will be
`small aberrations in otherwise consistent classification, such as a small group of mislabeled
`pixels surrounded by properly classified pixels. We use two algorithms during a postprocessing
`stage to help rectify these aberrations. To help remove small groups of misclassified pixels, we
`apply a process which “unclassifies” each pixel which was originally classified with a certainty
`below a set threshold. A second algorithm, known as “K-nearest neighbor” (KNN), reassigns the
`unclassified pixels to the region which is most popular amongst its neighboring pixels.
`Because objects in video are likely to reoccur in several shots and the system is run on shots
`individually, at this stage the author may wish to adjust region labels which have been set locally
`to values which represent a region or object globally. This could also mean adjusting mask
`values to a special number (i.e. 0) which corresponds to regions for which there is no action or
`link desired.
`At this stage the system has generated a second video sequence, which contains the segmentation
`information for every frame of the input sequence.
`Associating Actions with Objects
`If pixel-level identification of objects is possible (as it is with our system), a person can interact
`with the video during playback by indicating one or more pixels which comprise a particular
`object; the playback system can then cause a corresponding event or action to occur. This event
`could be the display of additional information about the object, such as the rendering of text or
`an image, or perhaps retrieving information from a URL on the World Wide Web. An event
`could also be a change in the way the video is presented, such as showing scenes from alternate
`camera angles or playing back a different part of the video which is relevant to the selected
`object. For example, the hyperlinked video “HyperCafe” [4] allows a person to navigate through
`various video clips of conversations occurring in a cafe by selecting actors in the video or icons
`which appear at certain times during the presentation.
`One possible implementation for enabling this association is to provide a database in which
`object labels can be cross-referenced with descriptions of the objects (if one of the desired
`actions is to render object information) or instructions for executing actions. Many of the
`existing hyperlinked video systems utilize this infrastructure; for example, VisualSHOCK
`maintains an “anchor list,” accessible via a drop-down menu system, where data structures
`containing link information are stored and referenced. The playback system for HyperSoap,
`built in a scripting environment called Isis [5], maintains an internal database consisting of a list
`of lists, in which each list corresponds to a particular object and contains text strings which
`include the name of the object, as well as actions associated with the object.
`Interacting with Hyperlinked Video
`
`Hulu
`Exhibit 1011
`Page 0004
`
`

`

`In this section, we describe the various issues which confronted us when we designed our
`playback system. Important user interface issues, including methods in which the presence of
`links are indicated in video, mechanisms used by the viewer to select objects, and approaches to
`initiating link events are described in detail. In many cases, our solutions to these issues vary
`greatly from existing implementations.
`One reason for this difference is that we focused our attention on user interaction mechanisms
`apart from the desktop PC and instead within the television paradigm. In some ways using a
`mouse on a PC may be natural for hyperlinked video, considering that people are used to using a
`mouse to activate hyperlinks on the World Wide Web. However, the desktop is not an ideal
`place to view video; additionally, many of the genres of content suitable for hyperlinked video
`are programs which people are accustomed to watching in their living rooms.
`We designed our system with the intent of simulating a future television viewing scenario, one in
`which a device with significant computational ability (e.g. a set-top box, enhanced television
`receiver or DVD player) could render video content while receiving directions about playback
`from both the viewer and the content itself. For our implementation we used a projection screen
`and a video projector which displayed video data transmitted by a workstation running Isis, a
`scripting environment which we used to create our playback software.
`When displaying hyperlinked video, we feel it is important to indicate the presence of hyperlinks
`to viewers in a manner which does not distract the viewer from the content of the video. This
`also enables a scenario in which the viewer chooses to watch the presentation in a passive
`manner, as one might read a document on the World Wide Web without choosing to click on
`hyperlinked words. Ideally, viewers might be able to adjust the manner in which hyperlink
`opportunities are presented to suit their own preferences.
`Many of the existing hyperlinked video playback tools utilize a common and minimally
`distracting method for presenting hyperlink opportunities to the viewer. Since the viewer will be
`interacting with the video using a pointing device (a mouse or its equivalent), when the viewer
`moves the cursor over a hyperlinked region, the cursor changes shape. At the same time, a name
`or description for the link is displayed in an area of the screen. IBM’s HotVideo software goes a
`step further and changes the cursor to an icon which corresponds to the type of information
`contained in a link, such as a small document if the link is text or a film reel if the object links to
`another video stream. HotVideo also displays an indicator at all times which changes color to
`indicate when there are hyperlink opportunities within the frames of video that are being shown
`to
`the viewer.
` This method
`is
`also used by WebTV’s WebPIP
`(http://developer.webtv.net/docs/itv/Default.htm) as there is always only one hyperlink
`opportunity in a video segment.
`Other approaches can be more distracting. Whenever a frame contains hyperlinked objects,
`HotVideo can also render the wireframe shapes which the author used to indicate the regions.
`This method is suitable when there are only a small number of objects present in each frame of
`video, but can cause the video to become quite cluttered for higher numbers of objects. It can
`also be confusing to the viewer if the linked object occupies a large portion of the video frame
`and may surround other objects, such as a wall in an interior shot of a house. Another potentially
`distracting display method which a HotVideo author can choose is to render the pixels within the
`wireframe shape in a different manner than the ones outside the shape, such as adjusting the
`brightness or color tint of those pixels. Many of these systems do not allow the viewer to choose
`the method in which all links are indicated, in favor of having the author determine this method
`per link.
`
`Hulu
`Exhibit 1011
`Page 0005
`
`

`

`Since our method allows for identifying objects at a pixel level, we are able to show the user the
`presence of hyperlinks by highlighting all of the pixels which comprise a selected object. This
`highlighting effect is achieved by using Isis’ ability to overlay an image buffer with a video
`stream in real time. During playback, there is an image buffer whose pixels are all set to a solid
`color. When the viewer selects a pixel which corresponds to a hyperlinked object, the playback
`script retrieves the segmentation mask for the next frame and sets the alpha channel of the pixels
`in the image buffer to correspond to the selected object’s pixels in the segmentation mask. In
`this manner, the only pixels in the image buffer which will be rendered are the ones which
`correspond to the selected object. Then, when the next frame of video is ready to be shown, Isis
`composites this frame of video with the image buffer (keeping the image buffer slightly
`transparent) thus indicating the presence of a hyperlink. If the viewer continues to select pixels
`from the same object, the playback script updates the alpha channel of the pixels in the image
`buffer. When the user stops selecting pixels, the script continues to update and render the image
`buffer, but adjusts the transparency of the buffer over time. This gives the viewer the impression
`that the link is “fading” out of the picture. We have also implemented two of these buffers
`simultaneously, which allows the system to show the presence of another object while the
`previously selected object may still be fading.
`The means by which the viewer of hyperlinked video selects links should be simple to use and
`feel as natural to the viewer as possible. Since people are comfortable using a remote control
`device to browse television programming, it makes sense for them to use this device to browse
`through hyperlinks in that content. WebTV provides buttons on the remote which can be pressed
`when the icon appears indicating the presence of a link for the present segment. The viewer can
`then choose to view the web page content referenced by the link or archive the link for later
`viewing. When several links are present, we envision a button on the remote control which
`would allow the viewer to cycle through the hyperlinked objects on the screen while a graphical
`overlay would be provided to indicate the location of each object as it is selected during the
`cycle. The remote might also incorporate a position touch pad, a joystick, or an inertial sensor
`For our demonstrations, we used a hand-held laser pointer as the interaction mechanism and set
`up a projection device which could both project frames of video and track the position of the
`laser on the screen. To do this, we added a small digital camera to a regular video projector and
`built the functionality into the Isis playback script to use the camera’s output. Since the laser
`pointer forms a red dot with significantly greater intensity than any of the red pixels of the
`projected video, the system looks at only the red channel of the camera’s output and applies a
`bounding-box algorithm to determine the center of the laser spot. We then apply a linear
`transformation to correct for possible keystone distortion, and compute a calibrated (x,y) position
`for the laser in units of screen pixels.
`Since the viewer may have no prior knowledge of which objects in the video are hyperlinked,
`there must be some method to allow the user to browse through objects shown on the screen
`before selecting one. To achieve this with our laser pointer interface, our playback script keeps
`track of the number of frames in which the viewer has pointed to the same object. If the viewer
`selects the object for more than a set number of frames, the script assumes the viewer has
`intended to select that object. To show the user that the object has been selected, the pixel values
`in the image buffer for highlighting are set to a solid color, thus modifying the appearance of the
`highlighting.
`The temporal nature of video raises an interesting issue regarding the display and timing of
`events associated with embedded hyperlinks. Whereas hyperlinked documents on the World
`Wide Web are stationary and can be read in any desired fashion, there is a flow to video content
`
`Hulu
`Exhibit 1011
`Page 0006
`
`

`

`which is generally not controlled by the viewer. While all of the existing systems direct web
`browsers to video links immediately, we have found that displaying lengthy or detailed
`information while the video continues to run may cause the viewer to feel distracted from the
`video content by having to read or examine an image simultaneously. Likewise, a delay in
`showing the link information may blur the context of the information such that it is no longer
`relevant to the video presentation. Thus events triggered by selecting hyperlinks must be
`presented in a way which maintains the continuity of the video while still presenting the desired
`link information to the viewer in a timely and meaningful fashion.
`Additionally, the system architecture and the capabilities of the viewer’s equipment may affect
`the degrees of freedom available. If the video is playing on-demand from a server, or locally
`from a DVD, random access and pausing are possible, and it’s not necessary to do much caching
`of content. The segmentation mask doesn’t necessarily even need to be transmitted, or it could
`be transmitted only when the user is actually pointing at something on the screen. In a broadcast
`environment, different content design might be needed, since pausing isn’t possible; also the
`display device needs the segmentation mask at all times. We’ve experimented with strategies
`appropriate for each case. In an early experiment, selecting a linked caused the video to be
`interrupted immediately to show the viewer the link content. After a set period of time, the link
`content would disappear and the video would resume from the last frame shown before the
`interruption. While some users enjoyed the instant response from the system, others found it
`somewhat troubling as their interactions would interrupt important moments within the content,
`such as cutting an actor off while they were delivering a line. Also, people tended to forget the
`context of the video content before the interruption and found themselves somewhat lost in the
`plot of the video when it resumed.
`Our next implementation yielded a pleasanter manner of interaction: once a viewer selected a
`hyperlink, the script would wait until the video reached a natural breakpoint (for instance, after
`an actor finishes speaking a set of lines) before interrupting and showing the link’s information.
`This yielded a much more favorable reaction from users of the system, although the initial
`reaction of many first-time users to waiting for information was that the playback mechanism
`was not working properly. Others who felt more interested in the link information than in the
`video content expressed that they would have preferred instant feedback.
`In addition to this mode of interaction, we decided to implement two additional modes for our
`system and gave the user the capability to toggle between all three modes during playback. The
`first additional mode overlaid an abridged version of the link information while the video
`continued to run. Based on the area of the screen where the user selected an object, the system
`would try to render the information such that it would be less likely to occlude other hyperlinked
`objects. This could be combined with a feature similar to WebTV’s such that each link was
`archived to enable the viewer to see the complete information after the end of the program.
`
`Hulu
`Exhibit 1011
`Page 0007
`
`

`

`Figure 1: A video frame with training points overlaid on it (top). The resulting segmentation
`(bottom).
`HyperSoap
`We used our authoring and playback systems to create HyperSoap, a four-minute prototype
`hyperlinked video drama whose content was closely patterned after daytime soap operas. We
`oversaw every aspect of the production of HyperSoap content, including scriptwriting,
`storyboarding and directing the shoot and postprocessing the video and audio. Every item on the
`set, including clothes, props and furnishings, was taken from the JCPenney catalog and was “for
`sale.” Pointing the laser-equipped remote at the screen highlighted selectable objects, and
`keeping the pointer fixed on one object for an extended period selected it. In instant-feedback
`mode (simulating a broadcast), a small text window would be displayed on the screen containing
`the item's description, brand name, price and a logo for the distributor. In breakpoint mode,
`more descriptive text boxes would be shown, as well as a still image resembling a photo of the
`item from a catalog. This same information could also be rendered at the end of the scene.
`
`Hulu
`Exhibit 1011
`Page 0008
`
`

`

`Figure 2: A viewer indicates a desired object by pointing at it with a special remote control.
`
`Figure 3: Information called up by selecting an object.
`In breakpoint mode, we used music to help maintain continuity between the video and the
`hyperlink information. A soundtrack was composed for HyperSoap with the unique quality that
`it could continue despite interruptions and still help to underscore the action being played out in
`the video. This was achieved by having the composer write pieces of music for specific parts of
`the scene which could be seamlessly looped by the playback script. When the scene changed
`from one part to the next, the music for the previous part would fade out while the new one
`began. As a result, when the user selected a link, the background music for the scene would
`continue through the interruption. Many users commented that the music helped to bridge the
`interruptions of the program content when objects were selected.
`Our authoring system produced good enough pixel-level segmentation of the 40 objects which
`appear in HyperSoap to permit the use of the segmentation mask for highlighting. There were a
`total of 45 shots which needed to be run through the system individually, where the number of
`
`Hulu
`Exhibit 1011
`Page 0009
`
`

`

`linkable objects per shot ranged from 10 to 20. We found that to maintain a consistently high
`quality of segmentation, we needed to retrain the system after approximately 30 frames, or
`approximately one second of video. This was largely done to prevent the propagation of errors
`during the tracking of the author-specified pixels. However, this still provided an exceptional
`level of automation of the segmentation process; typically an average of 5000 pixels were
`classified in each training frame by the author, meaning that for each 30-frame sequence the
`authoring system was required to classify 99.8% of the total pixels. Additionally, we found we
`were able to reduce the amount of data in the segmentation mask by subsampling the data
`vertically and horizontally by a factor of 2. This had a minimal effect on the quality of the
`upsampled mask when it is rendered.
`HyperSoap raised several interesting issues regarding the differences between traditional
`television programming and hyperlinked video content. A traditional television soap opera is
`designed in such a way that simply adding hyperlinks and allowing interruptions would likely
`detract from its effectiveness because they would obstruct the flow of one of the most important
`components of this format, the story.
`In designing the content for HyperSoap, we had to move away from a form where the story is
`central and any product placement or viewer interaction are decorations, to a new form where all
`of these components are equally important and dependent on each other for the survival of the
`presentation. The knowledge that everything would be for sale and that viewers would be
`interacting often was part of the program’s design from the very beginning. This entailed several
`changes in the way the scene was scripted, shot, and edited when compared to a traditional TV
`program.
`For example, shots were designed to emphasize the products for sale while not drawing too much
`attention away from the flow of the story. This often required closeups of certain objects which
`would have otherwise been difficult to see or select with our interaction device; actions needed
`to be written in the script so these closeups would not seem unnatural. Similarly, the scene was
`edited so that shots were either long enough or returned to often enough for viewers to spot
`products and decide to select them. Also, the story itself involved the products in ways that
`might increase their value to the consumer. Integrating the story with the products made
`interruptions to show product information less jarring overall.
`Summary and Applications
`We found that the knowledge that all of the objects in HyperSoap were for sale and that the
`program was designed for the dual purpose of entertainment and shopping motivated viewers to
`“stay tuned” and to interact with the products. As an additional way to motivate viewers to select
`products, we also implemented another version of HyperSoap in which selecting objects revealed
`details about the plot and characters in the drama which pertained to those items. We feel that
`both of these types of programming would create an intriguing form of interaction, were they to
`be implemented in a broadcast model. Other potential types of content which could benefit from
`“hyper” capabilities include how-to videos, particularly ones which involve complicated
`components, such as repairing an automobile or learning to fly an airplane. Likewise, training
`videos which demonstrate complex processes in which several people perform tasks
`simultaneously (e.g. a surgical procedure or a football play) could allow viewers with different
`backgrounds to learn more about the tasks or people whose purpose particularly interests them.
`This same idea could be applied to documentaries and educational programming, for example
`enabling children to learn more about particular aspects of nature by selecting animals and plants
`observed in “safari” footage, as if they were collecting specimens on a real safari.
`
`Hulu
`Exhibit 1011
`Page 0010
`
`

`

`Although we used the Isis language for our production, it is quite possible to use more standard
`tools (e.g. Java). While the computational demands of segmentation are quite high, display is
`well within the capabilities of a typical PC or a set-top box with graphical overlay capabilities.
`Acknowledgments
`We would like to thank Michael Ponder, JCPenney, and the JCPenney Headquarters Production
`Staff for their invaluable assistance with the HyperSoap shoot. We would also like to thank
`Kevin Brooks, Paul Nemirovsky and Alex Westner for providing their scriptwriting, scoring and
`Foley sound expertise. Also a special thank you to Julia Ann Lewis, Matthew Rupe and Andrew
`T. Chandler, for their tremendous acting talents. This work has been supported by the Digital
`Life Consortium at the MIT Media Laboratory.
`References
`[1] J. R. Smith and S.-F. Chang, “Local Color and Texture Extraction in Spatial Query,” IEEE
`Proc. Int. Conf. Image Processing, pp. III-1011—III-1016, 1996.
`[2] E. Chalom and V. M. Bove, Jr., “Segmentation of an Image Sequence Using Multi-
`Dimensional Image Attributes,” ,” IEEE Proc. Int. Conf. Image Processing, pp. II-525—II-528,
`1996.
`[3] E. Chalom, “Statistical Image Sequence Segmentation Using Multidimensional Attributes,”
`PhD thesis, MIT, Cambri