Multiple view wide-baseline trimap propagation for natural video matting

摘要

This paper presents a method to estimate alpha mattesudfor video sequences of the same foreground scene fromudwide-baseline views given sparse key-frame trimaps in audsingle view. A statistical inference framework is introducedudfor spatio-temporal propagation of high-confidence trimapudlabels between video sequences without a requirementudfor correspondence or camera calibration and motionudestimation. Multiple view trimap propagation integratesudappearance information between views and over timeudto achieve robust labelling in the presence of shadows,udchanges in appearance with view point and overlap betweenudforeground and background appearance. Results demonstrateudthat trimaps are sufficiently accurate to allow high-qualityudvideo matting using existing single view natural imageudmatting algorithms. Quantitative evaluation againstudground-truth demonstrates that the approach achievesudaccurate matte estimation for camera views separated byudup to 180◦ud, with the same amount of manual interactionudrequired for conventional single view video matting.