M F A T s o m c c p p o im w e u o c te 1 tr G p b b e Re Minglu Memorial U. of St. John’s, N gong@cs Figure 1: Real-tim ABSTRACT This paper presen sequences in real of novel Poiss multichannel co captured. A sim compute an init processing speed parallel processin online and auton mplemented to s which guides t evaluation on sti using the present obtained using comparable to th echniques. 1 INTRODUC Matting studie ransparency inf Generally speak problem: Given a color background obje background colo equation is satisf eal-Time V un Gong Newfoundland NL, Canada s.mun.ca (a) me video matting nts a novel matti l-time and onlin son equations olor vectors, as mple yet effectiv tial alpha matte d is achieved thro ng on the GPUs nomously, a mod separate foregro the automatic ill images show ted algorithm is m the global Pois hat of other state CTION es how to extract ormation from king, it tries to r image ۷ , whic cts, calculate the or ۰ , so that t fied: ۷= ߙ۴+ሺ1 Video Ma d Lia Univ. Lexing lwangd g results on the v using bilayer seg ing algorithm for ne. The algorithm that are deriv well as the d ve approach is e in the color ough optimizing . To process live dified backgroun und and backgro trimap generat ws that the alpha much more accu sson matting a e-of-the-art offli t foreground obje still images or o solve the fo ch contains both e matte ߙ, foregr the following a 1 − ߙሻ۰ atting usi ang Wang of Kentucky ton, KY, USA d@cs.uky.edu (b) video sequence c gmentation; (c) e r processing vid m is based on a s ved for handlin depth informatio also proposed space. Real-tim g the algorithm f e video sequenc nd cut algorithm ound, the result tion. Quantitativ a mattes extract urate than the on algorithm and a ine image mattin ects with per-pix video sequence ollowing ill-pos h foreground an round color ۴, an alpha compositin ( ng Multic Un Lex rya captured by Chua estimated alpha m eo set ng on to me for ces is of ve ed nes are ng xel es. ed nd nd ng (1) A va for still algorith 10-12, so far achieve devices video m On a separat segmen can ext powerf Howev bounda or even applied the stro handing in Figu matte in A re achieve only co qualitie When scene d to furth 2 R Here Reader matting 2.1 Vide Bayesia channel P Ruigang Yan niv. of Kentuck xington, KY, US ang@cs.uky.ed (c) ang et al. [3]: (a) matte; (d) compo ariety of techniqu l image matting hms are also prop 16]. However, d real-time video ed under studio s [12]. To the be matting based on another front, i ting foreground ntation [5, 15]. U tract the moving ful technique for ver, bilayer se aries surrounding n motion blur. Al d to alleviate the ong constraint us g objects with c ure 1. Applying n fuzzy areas is eal-time matting e this goal. The olor cue at real-ti es comparable to additional infor depth, is availabl her improve the m RELATED WOR e we limit the rs are referred t g techniques. Video Mattin eo matting is p an video mattin Poisson E g ky SA du Ed y ) input frame; (b) osite result. ues have been d g [1, 4, 6, 7, 9, posed for handli due to the high c matting for dy o settings using st of our knowle n color observati impressive resul d objects from Using just color i g foreground obj r video conferen egmentation ca g the foreground lthough the bord e aliasing proble sed in border ma omplex alpha m g real-time vide the next logical g algorithm is algorithm is abl ime speed and to o existing offline rmation, such a le, the algorithm matting accuracy RK discussion to o [17] for a nic g pioneered by C g approach, use Equations Yee-Hong Y Univ. of Albe dmonton, AB, C yang@cs.ualbe (d) automatic gener developed in the 13, 14, 18, 19] ing video sequen omputational co ynamic scenes c specially desig edge, technique f on only is not ye lts have been r live videos us information, the ect in real-time, ncing and live br annot capture d object caused der matting techn em along object atting limits its c matte, such as the eo matting to ex step. presented in th le to perform m o produce alpha e image matting as background c m can utilize the y. video matting ce survey on re Chuang et al. [3 ers are required t s Yang erta Canada erta.ca rated trimap past decade and several nces [1, 3, 8, ost involved, can only be gned optical for real-time et available. reported for sing bilayer eir algorithm making it a roadcasting. the fuzzy by hair, fur, nique [13] is boundaries, capability of e one shown xtract alpha his paper to matting using mattes with approaches. color and/or information approaches. lated image 3]. In their to manually 89 2010 Graphics Interface Conference 2010 31 May - 2 June, Ottawa, Ontario, Canada Copyright held by authors. Permission granted to CHCCS/SCDHM to publish in print form, and ACM to publish electronically.