Views
3 years ago

Vectorizing Cartoon Animations - Graphics & Geometric Computing ...

Vectorizing Cartoon Animations - Graphics & Geometric Computing ...

ZHANG ET AL.:

ZHANG ET AL.: VECTORIZING CARTOON ANIMATIONS 625Fig. 8. Vectorization results. Columns 1, 3: original video frames; columns 2, 4: corresponding vectorization results. (a)-(b) Two clips from Tom &Jerry; (c) A clip from Saint Seiya.Fig. 9. (a) Vectorization results using VectorMagic, our method and Adobe Illustrator Live Trace. (b) Region boundaries.region. We also provide an optional interactive step duringbackground reconstruction: the user can browse the perframe segmentation results and add desired regions to thebackground or remove regions from it. Our segmentationand interframe region matching algorithms are particularlyefficient, as they take advantage of the relatively simplenature of cartoons, allowing stable detection of large regions.On 640 480 input, segmentation typically takes 2-3 secondsper frame, background and foreground extraction take 4-8 seconds, and vectorization takes under 1 second; times varywith complexity of scene, but around 10 seconds per frame istypical, which means that an entire cartoon can be vectorizedin an acceptable length of time. In contrast, Ardeco [1] takes72-125 seconds per frame for 512 512 images, the StanfordVectorMagic image result in Fig. 9 took over 100 seconds, andthe method in [2] takes around 250-1,000 seconds per object.Fig. 9 compares our vectorization results for a singleframe to those produced by Stanford VectorMagic andAdobe Illustrator Live Trace. Our result gives more avisually appealing segmentation with fewer regions, withdecorative lines also identified.Figs. 10a and 10b show how our results can be readilyedited using Adobe Flash. The left-hand example shows,how Tom the cat has been extracted from one clip (seeFig. 8a), and inserted between the background and foreground(Jerry the mouse, in bed) of another clip, usingAuthorized licensed use limited to: Tsinghua University Library. Downloaded on March 07,2010 at 02:33:26 EST from IEEE Xplore. Restrictions apply.

626 IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 15, NO. 4, JULY/AUGUST 2009Fig. 10. Editing: (a) Composition of two clips taking a character from one scene and putting him in another. (b) Object editing changing the shape ofthe character’s ears.simple drag-and-drop. The right-hand example shows asimple edit to Jerry’s ear which was done on just two frames,and continues as the region is propagated through the video(the original image sequence is illustrated in Fig. 8b).(Editing was needed on two frames as Jerry’s ear underwenta large change in shape part way through the original inputsequence-note that the original was hand drawn.)Table 1 compares the file sizes of our vectorized outputto the original DivX compressed raster videos. The tablereports DivX 6 compressed video sizes (DivX), high-qualityvectorized output with decomposition detail level 0 andcurve fitting tolerance 4 (Vec1), and acceptable quality withdecomposition detail level 200 and curve fitting tolerance 20(Vec2). For high-quality output, our vectorization methodtypically leads to a file size of roughly half that required byDivX 6 compressed video. This can be further reduced byanother factor of 2 for lower but still acceptable qualityoutput. We note that if our method is used to acceleratenetwork transmission of cartoons, a desired bit rate can beadaptively achieved by adjusting the curve fitting and finedetail tolerances mentioned earlier. See also Fig. 13.7 DISCUSSION7.1 ComparisonWe now compare our results with those from the methods in[1], [2], [3], [4], [5], using images taken from their work (seeFig. 11). Sykora et al.’s method [3], [4], [5] imposes verystrong constraints on the cartoons processed, i.e. regionsmust have thick, closed black outlines and constant color. Asa result, their method failed on most cartoons we tried,making a direct comparison difficult. Our trapped-ballmodel for segmentation means that we can handle regionboundaries which need be neither closed nor clear, greatlyincreasing the applicability of our technique. Compared toArdeco [1], we can produce almost the same quality ofresults even for noncartoon images, despite these not beingour main goal. In Fig. 11b, the mean pixel difference betweenTABLE 1File Size Comparison (See Text)output of this method and the original image (L 2 norm inRGB color space) is 9 units, whereas in our result, the meanerror is 4 units. The resulting vectorized output from Ardecohas a more posterized appearance than our result: inparticular, the face and shoulder of the woman have morenoticeable steps in shading. Price and Barrett [23] and Sunet al. [2] give methods for vectorizing images that containobjects with complex but smoothly changing textures; theyuse meshes to describe those textures. Their work thus hasvery different motivation from ours: cartoons often containuntextured regions with generally well-defined boundaries.(Considering texture as a particular kind of color modelwould be an interesting extension to our current work.) Bothof these methods need users to manually create initialmeshes for each object in the image, and would be very hardto extend to video. However, these methods provide verylow reconstruction errors, often of less than 1 unit per pixel,which our method cannot. We note that, often, input videoclips suffer from compression artifacts. Our output mayactually have higher visual quality than the input: leastsquaresreconstruction errors are not a satisfactory metric forassessing vectorization quality.7.2 LimitationsOur method targets cartoons, which have a particular typeof artistic content, including smoothly (but not necessarilyuniformly) colored regions and decorative lines. Ourmethod produces suboptimal results when. Adjacent regions have similar color or edges whichare weak for some reason: boundaries may beincorrectly placed or regions incorrectly fused. Forexample, in Fig. 11c, the shadow of the pepper in ourresult is poorly segmented); while in Fig. 12b, poorsegmentation results where light crosses the character’shair.. Foreground objects adjacent to the background havea similar color to it. This can result in foregroundobjects being assigned to the background. Anexample is Jerry’s foot in Fig. 8b. (Colors arecompared after color modeling. After color modeling,the region color of the foot is considered to besimilar to that of the background.). There are complex textures in the scenes such asgrass or forest. In this case, an inappropriate colormodel is used for the kinds of regions present, andlarge colored regions do not represent the texturewell, resulting in large reconstruction errors.Authorized licensed use limited to: Tsinghua University Library. Downloaded on March 07,2010 at 02:33:26 EST from IEEE Xplore. Restrictions apply.

Introduction to OpenCL - Parallel Computing for Graphics
Mixing Graphics and Compute with multiple GPUs, Alina Alt
Mixing Graphics and Compute with multiple GPUs, Alina Alt - Nvidia
PDF 6.1M - Graphics & Geometric Computing Group
Computer Animation - Computer Graphics Laboratory at UCSD
Bitmap and Vector Graphics - Department of Computer Science
Vector Geometry for Computer Graphics
Generalized Displacement Maps - Graphics & Geometric Computing ...
Download PDF file - Computer Graphics and Geometric Modeling ...
Image Super-Resolution by Vectorizing Edges - Computer Graphics ...
Geometric Data Structures for Computer Graphics - Gabriel Zachmann
Computer Graphics and Geometric Ornamental Design - GRAIL
Cubic Mean Value Coordinates - Graphics & Geometric Computing ...
Vector Geometry for Computer Graphics - The University of Kansas
Geometric Data Structures for Computer Graphics Geometric Data ...
Geometric Modeling - Computer Graphics Laboratory
progressive 3d animated model - Computer Graphics Laboratory
Acrobat - Geometric Algorithms for Modeling, Motion, and Animation ...
Geometric Modeling - Computer Graphics Laboratory
Computer Graphics Animation for Highway Reconstruction after ...
computer-aided geometric design and computer graphics - OCW ...
Geometric modeling and computer graphics of kinematic ruled ...
Animation Model Simplifications - Computer Graphics Laboratory
Geometric fundamentals - Applied Geometry & Computer Graphics
Cloth Animation and Rendering - Computer Graphics Bonn