Views
2 years ago

Vectorizing Cartoon Animations - Graphics & Geometric Computing ...

Vectorizing Cartoon Animations - Graphics & Geometric Computing ...

ZHANG ET AL.:

ZHANG ET AL.: VECTORIZING CARTOON ANIMATIONS 619Fig. 2. Cartoon gallery.Fig. 1. (a) Original videos. (b) Vectorization results showing regionboundaries and decorative lines.provide a layered representation in which the regions anddecorative lines are vectorized; we also record theirmotions. Simple user-assistance is required to completethe background.The contributions of this paper are 1) the new trappedballsegmentation method, which is fast, supports nonuniformlycolored regions, and allows robust regionsegmentation even in the presence of imperfectly linkedregion edges; 2) the separate handling of decorative lines asspecial objects during image decomposition, avoidingresults containing multiple short, thin oversegmentedregions; and 3) extraction of a single patch-based backgroundfor all frames, which provides a basis for consistent,flicker-free animations.2 RELATED WORKSykora et al.’s [3], [4], [5] work on vectorization of cartoonanimations is the most closely related previous work toours. In their approach, cartoons must have a foregroundlayer consisting of dynamic regions each enclosed by clearlyvisible outlines. They rely heavily on correct and completedetection of the enclosing outlines, which are detected usingan edge detector similar to a Laplacian of Gaussian filter.Foreground and background are extracted using outlines ineach frame, and then a graph-based region matchingprocess is used to find the region relations and transformationsbetween frames. Due to this requirement for strongoutlines, their approach fails on many cartoons in practice,such as the one in Fig. 2. Our method can handle morecomplicated cartoons with nonuniform shading and weakeredges. Significantly, we are able to compute a high-qualitysegmentation without perfect edge detection.High-quality vectorization of cartoons requires accuratesegmentation into relatively few meaningful regions. Thereis a vast literature on image segmentation. Many sophisticatedcolor-based methods, such as mean-shift segmentation[6], typically generate an oversegmented result with toomany regions with flat shading and meaningless shapewhen applied to cartoons. Commercial software, such asAdobe Live Trace, CorelTrace, and AutoTrace, also typicallyproduces regions with flat shading. Ardeco [1] can findregions of quadratically varying color. However, as thismethod initially labels the pixels randomly and refines thelabeling, it also often produces many small regions, andhence, is unsuitable for our purpose. We generate largerinitial regions based on edge information and then refinethese regions using color compatibility to find preciseregion boundaries. Because we use edge information to findinitial regions, the final regions are larger and moremeaningful. However, we only label each pixel once, soour method is much faster than many other segmentationmethods, taking just a few seconds per frame.Qu et al. [7] proposed a colorization technique thatpropagates color over regions which is very suitable for“manga colorization,” A level-set-based method is used tosegment manga images. Since manga is drawn in black andwhite for printing, artists usually use patterns like hatchingand screening to illustrate shading, unlike cartoon animationswhich usually contain color regions with fewerpatterns. The only similarity is that both of our segmentationprocesses encounter the same problem of preventing regiongrowing through incomplete boundaries. Their methoddepends on tuning parameters to determine the size of thegaps to close, whereas our approach gives good segmentationresults without supervision, as explained in Section 4.Sun et al. also present a semiautomatic image vectorizationmethod [2]. Complex images must be manuallydecomposed into several semantic parts. Each part is thenmodeled using an optimized gradient mesh Coons patch,allowing for smooth color variation. Use of a gradient meshmeans that relatively few regions are needed to represent anobject. As the authors note, their method has problems withimages containing complicated topologies, very thin structures,or many small holes.Clearly, simply applying segmentation and vectorizationon a frame by frame basis will not produce good results,especially in the presence of raster compression artifacts andocclusion. Segmentation which is not coherent betweenframes will cause flickering in the vectorized output.Previous approaches to temporally coherent video segmentation[8], [9], [10] have tried to avoid such problems viaoptimization. Although such methods may be appropriatefor real-world video, they do not produce precise enoughresults for cartoons, as even minute inappropriate regionsand outlines are clearly visible, especially when present in thebackground, due to the smooth shading found in cartoons.We assume that the cartoon has an unchanging backgroundover a sequence of frames, allowing us to achieve temporalAuthorized licensed use limited to: Tsinghua University Library. Downloaded on March 07,2010 at 02:33:26 EST from IEEE Xplore. Restrictions apply.

620 IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 15, NO. 4, JULY/AUGUST 2009Fig. 3. Vectorization framework. The input at top left is the original cartoon video, and the output at bottom right the vectorized animation.coherence by extracting a unified background before detectionof foreground objects and their motions.Background subtraction has been extensively studied forcomplex scenes which, e.g., have dynamic backgrounds orchanging lighting [11], [12], [13]. Typically, two separateprocesses are used for background subtraction and foregroundregion extraction. While generally performing well,they often neglect pixels near boundaries, which may causeflickering of the background. As cartoons generally havefewer features with a clearer structure, we use a patchbasedbackground filling method at the same time asperforming foreground region extraction to ensure that allforeground pixels are appropriately allocated to regions.3 OVERVIEWTemporal video segmentation, which can segment a wholevideo into independent video sequences, each with adifferent (possibly moving) background shot, is a wellstudiedproblem [14], [15]. We assume that such segmentationhas already been performed. We focus on vectorizationof a single sequence, comprising a static background,possibly with camera motion relative to it, and foregroundmoving objects. Such static backgrounds are widely used incartoon making.Fig. 3 shows the framework of our system. We assumethat the input is a raster 2D animated cartoon, which is oflow quality due to lossy compression at some stage. Wevectorize each raster cartoon sequence as follows:. In each frame, decorative lines are detected first, andthese, together with edge detection results, are usedto build a mask. The image is then segmented intoregions using a trapped-ball method, controlled bythis mask.. To achieve interframe coherence, the frames in thesequence are registered by finding the homography,using the approach in [16]. A static panoramicbackground image is reconstructed by first initializingit with unchanging areas and refined by addingregions identified as belonging to the background.The moving objects are extracted as a foregroundlayer, together with their between-frame motions.. The background and foreground key objects arevectorized: their boundaries are represented ascurves and their interiors filled using quadratic (orpotentially any other) color models, and the vectorizedanimation is output.We now consider particular aspects in detail.4 SINGLE FRAME DECOMPOSITIONAn important requirement for improving visual coherenceis to decompose the cartoon image into relatively fewmeaningful objects. Typically, cartoon images contain twotypes of objects: colored regions and decorative lines-seeFig. 2. Colored regions need not have a uniform color, butmay be based on some simple model of color as a functionof position, e.g., a quadratic model.Due to the typically large differences in shading betweenneighboring regions in cartoons, edge information providesa strong hint for meaningful region segmentation. We thususe a Canny edge detector [17] to extract edge informationto guide image decomposition.However, directly using edge information to find regionboundaries and decorative lines has various challenges:. Especially when processing compressed video,whatever parameter settings are used, any simpleedge detector typically leads to various edges beingmissing, others containing gaps, and spuriousnoisy edges.. Edges found only to pixel-position accuracy areinsufficient for temporally coherent segmentation:camera motion does not generally involve a wholenumber of pixels per frame.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
Image Super-Resolution by Vectorizing Edges - Computer Graphics ...
Generalized Displacement Maps - Graphics & Geometric Computing ...
Download PDF file - Computer Graphics and Geometric Modeling ...
Geometric Modeling - Computer Graphics Laboratory
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 - Gabriel Zachmann
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 ...
computer-aided geometric design and computer graphics - OCW ...
Computer Graphics Animation for Highway Reconstruction after ...
Animation Model Simplifications - Computer Graphics Laboratory
Geometric modeling and computer graphics of kinematic ruled ...
Geometric Tools for Computer Graphics Problems list 1 Year 2012 ...
Realtime Performance-Based Facial Animation - Computer Graphics ...