Background Removal of Object Movies by Learning Shape Priors

環物影片技術已經被成功的使用在各種不同的應用中來拍攝和呈現物體。在把環物影片整合到所選擇的場景 (例如：環場) 前，有效且快速地將前景物體和背景場景分開是一項有挑戰性的工作。我們將這個問題稱之為環物影片去背。這篇論文提出一個新的環物影片去背方法，其中我們將物體的形狀資訊結合到去背的演算法當中。我們藉由使用體積最小剪割來做 3D 建模，並從當中學習到每一張影像所需要的形狀資訊。在這裡，為了改善建模的演算法，我們利用中軸來當約束。我們的去背方法只需要非常少的使用者介入，在自動的初始去背後，使用者只需選擇出可以接受的去背結果。和其他技術相比較，我們的方法不只提供了較好的去背結果，同時也有一個較好的 3D 建模結果。在環物影片去背的問題上，我們展示了令人印象深刻的改善。The object movie technique has been successfully used to capture and display 3D objects for manifold applications. Before integrating the object movies into the chosen scene source (e.g., a panorama), the challenge is how to efficiently and effectively separate the background from the interesting object. The problem will be referred to as object movie segmentation. This paper proposes a new object movie segmentation method that incorporates the shape prior into the segmentation algorithm. The shape prior introduced into every image of the object movie is learned from the 3D model reconstructed by the volumetric graph cuts algorithm. Here, the constraint derived from the discrete medial axis is used to improve the reconstruction algorithm. Our segmentation method requires only a small amount of user intervention, which is to select a subset of acceptable segmentations of the object movie after the initial segmentation process. Compared to other techniques, our method provides not only the better segmentation result but also the better 3D reconstruction result. Impressive improvement in object movie segmentation is demonstrated.1 Introduction 1 2 Related Works 5 3 Overview of Our Approach 8 4 Automatic Initial Segmentation 10 4.1 Graph Cut Image Segmentation 10 4.2 Trimap Labeling 11 4.2.1 B-labeling 12 4.2.2 F-labeling 13 5 Segmentation with Shape Priors 16 5.1 Volumetric Graph Cuts 16 5.2 Discrete Medial Axis Constraint 18 5.2.1 Energy Function Analysis 18 5.2.2 Imposing the DMA Constraint 20 5.3 Segmentation Refinement 22 6 Experiments 23 6.1 Initial Segmentation Results 24 6.2 Learning Shape Prior 25 6.3 Rectification of Segmentation Errors 28 7 Conclusion 31 Bibliography 3