`
`MICHAL IRANI, MEMBER, IEEE, AND P. ANANDAN, MEMBER, IEEE
`
`information. It provides visual
`Video is a rich source of
`information about scenes. This information is implicitly buried
`inside the raw video data, however, and is provided with the cost
`of very high temporal redundancy. While the standard sequential
`form of video storage is adequate for viewing in a “movie mode,”
`it fails to support rapid access to information of interest that is
`required in many of the emerging applications of video. This paper
`presents an approach for efficient access, use, and manipulation
`of video data. The video data are first transformed from their
`sequential and redundant frame-based representation, in which
`the information about the scene is distributed over many frames,
`to an explicit and compact scene-based representation, to which
`each frame can be directly related.
`the video data supports
`This compact reorganization of
`nonlinear browsing and efficient indexing to provide rapid access
`directly to information of interest. This paper describes a new
`set of methods for indexing into the video sequence based on the
`scene-based representation. These indexing methods are based
`on geometric and dynamic information contained in the video.
`These methods complement the more traditional “content-based
`indexing” methods, which utilize image-appearance information
`(namely, color and texture properties) but are considerably
`simpler to achieve and are highly computationally efficient.
`Keywords—Compact video representations, mosaics, video an-
`notation, video browsing, video compression, video data bases,
`video indexing, video manipulation.
`
`I.
`
`INTRODUCTION
`
`The emergence of video as data and a source of informa-
`tion on the computer opens the potential for new ways of
`accessing, viewing, and manipulating the contents of video.
`These include direct nonlinear access to video frames and
`sequences of interest, new modes of viewing that give the
`viewer control over how the video is viewed, annotation
`and manipulation of objects and scenes in the video, and
`merging of text and graphics with the video data.
`While the standard manner of representing video as a
`sequence of frames is adequate for viewing it in a movie
`mode,
`it does not support
`the type of interaction with
`video information described above. Currently, the only way
`
`Manuscript received July 20, 1997; revised November 30, 1997. The
`Guest Editor coordinating the review of this paper and approving it for
`publication was A. M. Tekalp.
`M. Irani was with the Sarnoff Corporation, Princeton, NJ 08540 USA.
`She is now with the Department of Applied Math and Computer Science,
`The Weizmann Institute of Science, Rehovot 76100 Israel.
`P. Anandan was with the Sarnoff Corporation, Princeton, NJ 08540
`USA. He is now with Microsoft Corporation, Redmond, WA 98052 USA.
`Publisher Item Identifier S 0018-9219(98)03562-2.
`
`to access the information of interest is by sequentially
`scanning the video. The only way to manipulate, annotate,
`or edit the video is by processing the video frame by frame.
`This process is both slow and tedious.
`This paper presents a new approach for efficient access,
`storage, and manipulation of video data. Our approach is
`based on the fact that a video sequence contains many views
`of the same scene taken over time, from either a moving or
`a stationary camera. Hence, the information that is common
`to all the frames is the scene itself. This information is dis-
`tributed over many frames, however, at the cost of very high
`temporal redundancy, and is found only implicitly in the
`video data. We transform the video data from a sequential
`frame-based representation, in which this common scene
`information is distributed over many frames, into a single
`common scene-based representation to which each frame
`can be directly related. This representation then allows
`direct and immediate access to the scene information,
`such as static locations and dynamically moving objects.
`It also eliminates the redundancy between the different
`views of the scene contained in the frames and results in
`a highly efficient and compact representation of the video
`information. Hence, the scene-based representation forms
`the basis for direct and efficient access to and manipulation
`of the video information and supports efficient storage and
`transmission of the video data.
`The scene representation is composed of three compo-
`nents.
`
`1) Extended spatial information: this captures the ap-
`pearance of the entire scene imaged in the video clip
`and is represented in the form of a few (often just one)
`panoramic mosaic images constructed by composing
`the information from the different views of the scene
`in the individual frames into a single image.
`
`2) Extended temporal information: this captures the mo-
`tion of independently moving objects in the scene
`(e.g., in the form of their trajectories).
`
`the three-
`this captures
`3) Geometric information:
`dimensional (3-D) scene structure, as well as the
`geometric transformations that are induced by the
`motion of the camera, and maps the frames to the
`common mosaic image.
`
`PROCEEDINGS OF THE IEEE, VOL. 86, NO. 5, MAY 1998
`
`905
`
`0018–9219/98$10.00 ª
`
`1998 IEEE
`
`Prime Focus Ex 1038-1
`Prime Focus v Legend3D
`IPR2016-01243
`
`
`
`Taken together, these three components provide a compact
`description of the video data.
`We construct the common scene-based representation by
`measuring and interpreting the image motion within the
`video clip. Regions of the video frames corresponding to
`the static and dynamic portions of the scene are determined.
`The geometric transformations and the 3-D scene structure
`are recovered as a part of this process. This process is done
`automatically, without any information about the camera
`calibration or the scene.
`Once the common scene-based representation is con-
`structed, it forms the basis for direct and efficient brows-
`ing, indexing, and manipulation of the video data. Brows-
`ing is done by skimming a collection of images that
`“summarize” the video data. We refer to these images
`as visual summaries. These summaries visually describe
`the video information in a compact and succinct fash-
`ion and can serve as a visual table of contents for the
`video.
`Since the mosaics capture the information that is common
`to all
`the frames,
`they provide the means directly to
`index into and manipulate the individual frames. Both the
`static and dynamic portions of the video sequence can
`be accessed this way. These indexing methods are based
`on geometric and dynamic information contained in the
`video. These complement the more traditional approach to
`“content-based indexing,” which utilizes image appearance
`information (namely, color and texture properties) [7], [9],
`[10], [26], but are considerably simpler to achieve and are
`computationally highly efficient. The existing appearance-
`based methods themselves can also be used more efficiently
`within the scene-based representation when applied directly
`to the mosaic image (i.e., to the appearance component
`of our representation), rather than to the individual video
`frames one by one.
`The rest of this paper is organized as follows. Section II
`presents the common and compact scene-based representa-
`tion, to which each frame is directly related. Section III
`explains how to use the scene-based representation to
`browse, index, and manipulate video data efficiently and
`rapidly. Section IV reviews the techniques used for con-
`structing the scene-based representation from raw video
`sequences. Section V concludes this paper.
`
`II. FROM FRAMES TO SCENES
`
`Video is a rich data source. It provides information about
`scenes. This information is buried inside the raw video
`data, however, and is provided at the cost of very high
`temporal redundancy (e.g., every scene point is displayed
`repeatedly in numerous consecutive frames). In this section,
`we first review the fundamental components of information
`in a video stream (Section II-A). Then we make use of
`these information components to transform the video from
`an implicit and redundant frame-based representation to an
`explicit and nonredundant scene-based representation that
`is common to all frames (Section II-B).
`
`A. The Three Fundamental Information
`Components of Video
`Video extends the imaging capabilities of a still camera
`in three ways. First, although the field of view of each
`single image frame may be small, the camera can be panned
`or otherwise moved around in order to cover an extended
`spatial area. However,
`the extended spatial information
`acquired by the video is not available in a coherent form.
`It is distributed among a sequence of frames and is hard
`to use.
`The second, and perhaps the most common, use of
`video is to record the evolution of events over time.
`Again, however, this extended temporal information is not
`explicitly represented but distributed over a sequence of
`video frames. While it is natural for a human to view it as
`a movie, this representation is not particularly suitable for
`analytic purposes.
`Third, a video camera can be moved in order to acquire
`views from a continuously varying set of vantage points.
`This induces image motion, which depends on the 3-D
`geometric layout of the scene and the motion of the camera.
`However, this geometric information is also only implicitly
`present and is not directly accessible from the standard
`sequential video representation.
`Thus, the total information contained in the video data
`consists of the three scene components mentioned above.
`However, this information is distributed among the frames
`and is implicitly encoded in terms of image motion. There-
`fore, a natural way to reorganize the video data is in terms
`of these three scene components. Moreover, such a reorga-
`nization removes the tremendous redundancy that is present
`in the source video data. This scene-based organization is
`highly efficient since it directly and uniquely maps onto the
`information in the scene. Therefore, it facilitates efficient
`interaction and manipulation and supports very efficient
`storage and transmission.
`
`B. The Scene-Based Representation
`To bring out the common scene information contained in
`the video, and make it more directly accessible, we first
`transform the video from its implicit and redundant frame-
`based representation to an explicit and compact scene-based
`representation. In this section, we introduce the scene-based
`representation. In Section IV, we elaborate on the details of
`the representation and explain how it is constructed from
`the video data.
`temporally segmented into
`The video stream is first
`scene segments, which are subsequences of the input video
`sequence. A beginning or an end of a scene segment
`is automatically detected wherever a scene cut or scene
`change occurs in the video. The scene cuts are characterized
`typically by drastic changes in the frame content, which are
`directly reflected in the distribution of color and the gray
`levels in the image, or in the image motion (e.g., see [9]
`and [37]). These changes are relatively simple to detect.
`Each scene segment
`is subsequently parsed into the
`three fundamental components of video (see Section II-A),
`namely, the static background scene, the dynamic moving
`
`906
`
`PROCEEDINGS OF THE IEEE, VOL. 86, NO. 5, MAY 1998
`
`Prime Focus Ex 1038-2
`Prime Focus v Legend3D
`IPR2016-01243
`
`
`
`objects, and the geometric information. These components
`are organized as described below.
`Corresponding to the three fundamental components, the
`scene-based representation is divided into three parts.
`1) Panoramic Mosaic Image: This captures an extended
`spatial view of the entire scene visible in the video clip in a
`single (or sometimes a few) “snapshot” image(s) (e.g., see
`Fig. 1). This image captures the appearance of the static
`portions of the scene.
`The mosaic image is constructed by first aligning all
`the frames with respect to the common coordinate system
`(which becomes also the mosaic coordinate system) and
`then integrating all these frames to form a single image.
`Different methods of integration can be employed (e.g.,
`temporal average, temporal median, superresolution, etc.).
`These are described in more detail in [12].
`The mosaic representation removes the redundancy con-
`tained in the overlap between successive frames and rep-
`resents each spatial point only once. Mosaics have been
`previously used as an effective way of creating panoramic
`views of a scene from video sequences [3], [16], [20],
`[23], [31], [32]. Until now, however, they have not been
`used as an information component within a scene-based
`representation, which provides direct and efficient access
`to video data.
`Section IV describes a hierarchy of mosaic representa-
`tions. The hierarchy corresponds to increasing complexity
`levels in the camera motion and in the 3-D scene structure.
`2) Geometric Transformations: These relate the different
`video frames to the mosaic coordinate system. The geo-
`metric transformations contain the information necessary
`to map the location of each scene point back and forth
`between the panoramic mosaic image(s) and the individual
`frames. Corresponding to the hierarchy of the panoramic
`mosaic representations, there exists a hierarchy of represen-
`tations of the geometric transformations. These range from
`global parametric two-dimensional (2-D) transformations to
`more complex 3-D transformations and are described in
`Section IV.
`3) Dynamic Information: This is the information about
`moving objects, which are not captured by the static
`panoramic mosaic image. Moving-object
`information is
`completely captured by representing the extended time
`trajectories of those objects as well as their appearance.
`Such a complete representation is needed, e.g.,
`for
`video compression (since the video frames need to be
`reconstructed from the scene representation). To access,
`browse, index, and annotate the video (as presented in
`Section III), however, the trajectory information alone is
`sufficient. The trajectory of the center of mass of each
`detected moving object (i.e., a single image point per
`moving object per frame) is maintained. These trajectories
`are represented in the coordinate system of the mosaic
`image, which is common to all
`the frames.
`In the
`common coordinate system, time continuity, continuous
`tracking, and the temporal behavior of
`the moving
`object can be analyzed more effectively (see Figs. 3 and
`5).
`
`the three components of our scene-based rep-
`Thus,
`resentation form a compact representation of the video
`clip. The compactness results from the fact
`that every
`scene point is presented only once in the mosaic image,
`while in the original video clip, it is observed in multiple
`frames. This compactness of the scene-based representation
`facilitates very high compression (and we have developed
`such algorithms for very-low-bit-rate compression [13]). In
`this paper, we focus on the power of this representation
`for video indexing and manipulation. Section III describes
`how this representation can be used for efficiently accessing
`and manipulating the video data. Section IV describes the
`methods for constructing the scene-based representation.
`
`III. FROM SCENES TO VISUAL SUMMARIES AND INDEXING
`
`Once a video sequence is transformed from the frame-
`based representation to the scene-based representation, it
`forms the basis for the user’s interaction with the video. The
`user can initially preview the video by browsing through
`visual summaries of the various video clips. These visual
`summaries can serve as a visual table of contents of the
`video data. When a scene of interest is detected by the
`user, he can either request to view only that portion of the
`video or further index into individual video frames. The
`detected frames of interest can then be either viewed or
`manipulated by the user.
`
`A. Visual Summaries—A Visual Table of Contents
`
`There are two types of visual summaries of video clips
`through which a user can browse. These are captured by
`two types of mosaic images, which are constructed from
`the video clip of a scene.
`1) The Static Background Mosaic: The video frames of a
`single video segment (clip) are aligned and integrated into
`a single mosaic image. This image provides an extended
`(panoramic) spatial view of the entire static background
`scene viewed in the clip in a single “snapshot” image
`and represents the scene better than any single frame.
`This image does not include any moving objects. The user
`can visually browse through the collection of such mosaic
`images to select a scene (clip) of interest.
`Figs. 1 and 2(b) display some examples of static back-
`ground mosaic images.
`2) The Synopsis Mosaic: While the static mosaic image
`effectively captures the background scene, it contains no
`representation of the dynamic events in the scene. To
`provide a summary of the events, we create a new type
`of mosaic called the synopsis mosaic. This is constructed
`by overlaying the trajectories of the moving objects on top
`of the background mosaic. This single “snapshot” image
`provides a visual summary of the entire dynamic foreground
`event that occurred in the video clip.
`Fig. 3 graphically illustrates the trajectory associated with
`a moving object in a synopsis mosaic.
`Fig. 2(c) provides a summary of the entire event in the
`baseball video clip.
`
`IRANI AND ANANDAN: VIDEO INDEXING BASED ON MOSAIC REPRESENTATIONS
`
`907
`
`Prime Focus Ex 1038-3
`Prime Focus v Legend3D
`IPR2016-01243
`
`
`
`(a)
`
`(b)
`
`Fig. 1. Static background mosaic of an airport video clip. (a) A few representative frames from the
`minute-long video clip. The video shows an airport being imaged from the air with a moving camera.
`The scene itself is static (i.e., no moving objects). (b) The static background mosaic image, which
`provides an extended view of the entire scene imaged by the camera in the one-minute video clip.
`
`To allow for comprehensive display of multiple tra-
`jectories (corresponding to multiple moving objects), the
`trajectory of each moving object is uniquely color coded.
`Figs. 4 and 5 provide visual summaries of airborne video
`clips each with multiple moving objects. Fig. 4 shows a
`flying airplane and a moving car on the road. Fig. 5 shows
`a flying airplane, three parachuters that were dropped from
`the plane, and a moving car.
`The natural mode of operation for the user is first to
`browse through the visual summary mosaics to identify a
`
`few scenes of interest. Once the user has identified a scene
`(i.e., mosaic) of interest, he proceeds to directly access
`and/or manipulate individual video frames associated with
`only a portion of the scene that is of interest to him. The
`scene-based representation supports this type of indexing.
`Two new types of indexing methods are presented: 1)
`indexing based on location (geometric) information, and
`2) indexing based on dynamic information. These are made
`possible directly via the geometric coordinate transforma-
`tions that relate the different frames to the mosaic image and
`
`908
`
`PROCEEDINGS OF THE IEEE, VOL. 86, NO. 5, MAY 1998
`
`Prime Focus Ex 1038-4
`Prime Focus v Legend3D
`IPR2016-01243
`
`
`
`(a)
`
`(b)
`
`(c)
`
`Fig. 2. Visual summaries of a baseball video clip. (a) A few representative frames from the video
`clip. The video shows two outfielders running, while the camera is panning to the left and zooming
`on the two baseball players. (b) The static background mosaic image, which provides an extended
`view of the entire scene captured by the camera in the video clip. The “missing” regions at the
`top left and bottom left were never imaged by the camera because at that point, it was zoomed on
`the two players (e.g., frame 80). (c) The synopsis mosaic, which provides a visual summary of the
`entire event. It shows the trajectories of the two outfielders in the context of the mosaic image.
`
`Fig. 3. Synopsis of a moving object. The trajectory of the moving object is depicted in the
`synopsis mosaic. This shows the motion of the moving object after cancellation of the background
`(camera-induced) motion. With each point on the trajectory is associated a frame number (i.e., the
`“time” when the moving object was at that location).
`
`through the moving objects information that was estimated
`in the formation of the mosaic-based scene representation
`(Section II-B). The access and manipulation of selected
`video frames is done directly from the mosaic-based visual
`summaries. These location and dynamic indexing methods
`complement
`the more traditional approach to “content-
`
`based indexing,” which utilizes image-appearance informa-
`tion (e.g., color and texture) [7], [9], [10], [26]. However,
`our methods are considerably simpler to achieve and are
`highly computationally efficient.
`The remainder of this section describes these modes of
`video indexing and manipulation.
`
`IRANI AND ANANDAN: VIDEO INDEXING BASED ON MOSAIC REPRESENTATIONS
`
`909
`
`Prime Focus Ex 1038-5
`Prime Focus v Legend3D
`IPR2016-01243
`
`
`
`(a)
`
`(b)
`Fig. 4. The visual summary of a flying plane video clip. (a) A few representative frames from
`the minute-long video clip. The video shows an airplane flying from right to left (during takeoff).
`A car driving on a road is visible for a few frames. (b) The synopsis mosaic, which provides
`a visual summary of the entire video clip, showing the trajectories of all moving objects in the
`context of the mosaic image. Each detected and tracked moving object is color coded uniquely
`(plane: green; car: yellow).
`
`B. Location (Geometric)-Based Indexing
`Once a few scenes of interest (in the form of visual
`summaries) have been selected, the user proceeds to access
`the video frames themselves. The user selects a scene point
`
`(or several points) in the mosaic image. The geometric
`coordinate transformations map the selected scene point(s)
`from the mosaic image to its location in the coordinate
`system of each of the video frames. All frames containing
`
`910
`
`PROCEEDINGS OF THE IEEE, VOL. 86, NO. 5, MAY 1998
`
`Prime Focus Ex 1038-6
`Prime Focus v Legend3D
`IPR2016-01243
`
`
`
`(a)
`
`(b)
`Fig. 5. The visual summary of a parachuters video clip. (a) A few representative frames from the
`30-second-long video clip. The video shows an airplane flying from left to right, dropping three
`parachuters. A car driving on a road is visible for a few frames. The parachuters are very small
`(tiny white dots) and difficult to see in a static image, but they are easily detectable in video, as they
`have different motion than the background. They are depicted in the synopsis mosaic by the green,
`red, and blue trajectories. In the video sequence, they become visible gradually, as their parachutes
`open—first the left parachuter, then the right one, and last the middle one. This becomes clearer
`in the annotated video displayed in Fig. 10.
`(b) The synopsis mosaic, which provides a visual
`summary of the entire video clip, showing the trajectories of all moving objects in the context of
`the mosaic image. Each detected and tracked moving object is color coded uniquely (parachuters:
`green, red, blue; car: purple; plane: yellow).
`
`inside their field of view are
`the selected scene point
`therefore instantaneously determined. The user can view
`the subsequence of the video that contains only the frames
`
`with the selected scene point (or points). When these frames
`are not consecutive in time (e.g., if the selected portion of
`the scene was revisited by the camera multiple times), then
`
`IRANI AND ANANDAN: VIDEO INDEXING BASED ON MOSAIC REPRESENTATIONS
`
`911
`
`Prime Focus Ex 1038-7
`Prime Focus v Legend3D
`IPR2016-01243
`
`
`
`Fig. 6. Location-based indexing. Selection of a scene point in the mosaic image generates a display
`of all frames whose field of view contains the selected scene point. These are frames i; j; and k.
`In the figure, these frames are displayed as a collection of frames, but in reality, they are displayed
`as a video sequence.
`
`multiple subsequences (corresponding to consecutive frame
`groups) are displayed to the user.
`Fig. 6 demonstrates an indexing process. Selection of a
`scene point in the mosaic image generates a display of
`all frames whose field of view contains the selected scene
`and
`. In the figure, these
`point. These are frames
`frames are displayed as a collection of frames, but in reality,
`they are displayed as a video sequence.
`In addition to manual scene-point selection, this repre-
`sentation also provides a basis for efficiently indexing into
`the video using existing automatic detection methods. For
`example, if a region is searched using an appearance-based
`detection method (e.g., template correlation or search based
`on color or texture attributes [7], [9], [10], [26]), then
`instead of applying these search methods individually to
`each frame, they can be applied just once to the common
`mosaic image. Once it is detected in the mosaic image, the
`location-based indexing mechanism can be used to retrieve
`the corresponding frames.
`1) Editing and Annotation: The compact mosaic repre-
`sentation can be used not only to access video frames
`but also to edit, annotate, and manipulate these frames.
`For example, the same mechanism used for indexing is
`also used to efficiently inherit annotations from the mosaic
`image onto scene locations in the video frames.
`The annotation is specified by the user just once on the
`mosaic image rather than tediously specifying it for each
`
`and every frame. This can be further extended to efficiently
`edit video clips by inserting or deleting an object in the
`mosaic image, hence inserting or deleting that object in all
`corresponding video frames.
`Fig. 7 graphically illustrates a video annotation process.
`Fig. 8 shows an example of annotating airborne video of
`an airport scene.
`
`C. Dynamic (Moving-Objects)-Based Indexing
`Since the synopsis mosaic provides a snapshot view
`of an entire dynamic event, it can be used for indexing
`based on temporal events. In the synopsis mosaic,
`the
`motion of an object is represented as a trajectory in the
`common coordinate system; hence, the temporal event has
`been transformed into a spatial representation. Marking a
`segment on the trajectory is thus equivalent to marking a
`time interval, which enables access and display of all frames
`in this time interval.
`More specifically, all frames containing a selected mov-
`ing object can be immediately determined and accessed,
`as can the location of the moving object in each of these
`frames. The user can select an object of interest whose track
`is marked on the synopsis mosaic. Since the trajectories
`of the moving objects in the mosaic coordinate system
`are precomputed (as well as which point on the trajectory
`corresponds to which frame), all frames containing that
`object are immediately accessed and viewed. The location
`
`912
`
`PROCEEDINGS OF THE IEEE, VOL. 86, NO. 5, MAY 1998
`
`Prime Focus Ex 1038-8
`Prime Focus v Legend3D
`IPR2016-01243
`
`
`
`Fig. 7. Location-based annotation. Annotation of a selected scene point in the mosaic image leads
`to automatic annotation of all relevant frames (i; j; and k) with the selected annotation and at the
`appropriate image coordinate, i.e., that which corresponds to the selected scene point in each of
`the frames.
`
`of that object in each frame is estimated through the basic
`geometric coordinate transformations (those that correspond
`to the camera-induced motion). In a similar manner, the
`moving objects in the video frames are efficiently annotated
`or manipulated by annotating the synopsis mosaic, without
`the need for the user to perform the operation repeatedly
`on a frame-by-frame basis.
`Fig. 9 shows an example of annotating moving objects
`using the plane video, whose synopsis mosaic was shown in
`Fig. 4. The figure displays the selected annotations on the
`synopsis mosaic. Representative output frames are shown,
`in which the annotations are automatically inherited from
`the mosaic. Note that the annotations “move” together with
`the moving objects.
`Fig. 10 shows an example of video annotation using the
`airborne parachuters video. The figure displays the selected
`annotations on top of the synopsis mosaic image. Both
`moving objects and stationary scene points are annotated.
`Representative frames from the automatically annotated
`video clip are also displayed. Note that annotations of
`moving objects “move” together with the moving objects,
`while annotations of static scene points (e.g., “building”)
`remain stationary with respect to the background scene
`(i.e., they preserve the background motion induced by the
`moving camera).
`Note also that estimating the trajectories of moving ob-
`jects in the common mosaic coordinate system allows more
`reliable detection and tracking of moving objects, even
`
`when they are very small (such as the three parachuters in
`Fig. 5). This is because a “temporal coherence” constraint
`can be used during moving-object detection and tracking
`after removal of the background motion. Assuming that
`object velocities do not change too rapidly, the detection
`of moving objects within each frame can be guided by the
`trajectory of the objects in a few previous frames. This
`leads to better separation between small moving objects
`and noise as well as enables recovery from losing an
`object for a few frames (e.g., due to occlusion or bad
`detection). The missing portion of the trajectory is smoothly
`interpolated/extrapolated from the neighboring frames.
`
`IV. BUILDING THE SCENE-BASED REPRESENTATION
`In Section II-B, we introduced the basic components of
`the scene-based representation. In this section, we provide
`the details of the scene-based representation (Section IV-
`A), followed by a review of the methods used for its
`construction (Sections IV-B and IV-C). This section serves
`mainly as a review of methods that have been previously
`published; these methods are briefly outlined here in order
`to make the paper self-contained.
`
`A. The Detailed Scene-Based Representation
`1) Panoramic View: The panoramic view of the scene is
`captured by one or several mosaic images. We present a
`hierarchy of such mosaic representations. The hierarchy
`
`IRANI AND ANANDAN: VIDEO INDEXING BASED ON MOSAIC REPRESENTATIONS
`
`913
`
`Prime Focus Ex 1038-9
`Prime Focus v Legend3D
`IPR2016-01243
`
`
`
`(a)
`
`(b)
`
`Fig. 8. Annotation of the airport video clip. (a) A stationary car is annotated once on the mosaic
`image (“car”). (b) A few representative frames from the video clip with the annotations inherited
`from the mosaic image. The annotations are incorporated into the video frames automatically and
`instantly through the geometric coordinate transformations that map each frame onto the mosaic
`image. Some video frames from the raw video clip are displayed in Fig. 1.
`
`corresponds to increasing complexity levels in the camera
`motion and in the 3-D scene structure.
`a) 2-D parametric mosaic image: The simplest repre-
`sentation is a mosaic image constructed by aligning all the
`frames to a single coordinate system using 2-D parametric
`coordinate transformations. We refer to such a mosaic
`as a 2-D parametric mosaic image. The cases when the
`camera-induced motion can be modeled as a 2-D parametric
`transformation can be divided broadly into three categories
`(see Section IV-B1): 1) when the translational motion of
`the camera is negligible, i.e., camera motion can be ap-
`proximated by only 3-D rotations and zooms, 2) when the
`scene is planar, or 3) when the 3-D scene is sufficiently
`distant from the camera such that it can be approximated
`by a nearly flat 2-D surface. We refer to these scenarios
`as 2-D scenes.
`The examples given in Section III belong to this class of
`scenarios. For example, the baseball sequence (Fig. 2) was
`
`captured by a panning camera (i.e., pure rotation), while the
`other sequences in that section (Figs. 1, 4, and 5) were taken
`by an airborne camera. Hence, the scene was sufficiently
`distant from the camera and could be well approximated
`by a flat 2-D surface.
`parallax representation: The next level of
`b) Plane
`complexity arises when the 3-D deviations from the 2-D
`planar surface approximation (when combined with the
`camera translation) result
`in measurable parallax image
`motion relative to the surface. In this case,
`the visual
`appearance of the scene is still captured by a mosaic image,
`as in the previous case, while the geometric component of
`the representation also encodes the 3-D parallax relative
`to the planar surface (see Section IV-B2). The parallax
`information is captured in the geometric component of the
`representation and is taken into account while combining
`the different frames into a single mosaic [12]. We refer to
`parallax representation.
`this representation as the plane
`
`914
`
`PROCEEDINGS OF THE IEEE, VOL. 86, NO. 5, MAY 1998
`
`Prime Focus Ex 1038-10
`Prime Focus v Legend3D
`IPR2016-01243
`
`
`
`(a)
`
`(b)
`Fig. 9. Annotation of the flying plane video clip. (a) The annotations are defined once on the
`synopsis