以動向平滑為基礎之影片分割

Abstract

本論文探討並實作了一個基於圖分割演算法的互動式影片分割系 統。近來,基於圖分割演算法的圖像分割、影片分割於電腦圖學與電 腦視覺研究界甚為普遍。然而,絕大多數的關連研究僅僅使用了影片 本身的色彩資訊,作為主要的分割依據。這在前景與背景有部分區域 在色彩上甚為相似的狀況下,容易產生錯誤。而不幸地,這樣的條件 並不罕見,特別是當拍攝對象並非在棚拍等人工環境之下拍攝,而是 以日常場景作為背景之時。因此,在本論文之中,我們提出了除了色 彩之外的依據進行影片分割的演算法。我們觀察到前景的動向與背景 經常是相當不同的,因此,選擇結合色彩以及動向資訊共同進行影片 分割。此外,本系統尚且擴充了原本採用於圖片分割領域的漸進式分 割,使其能夠用於影片分割。最後,我們將本系統的結果與關連研究 進行了比較,以實例證實了本系統的效能確實優於既往研究。

In this thesis, we present an interactive graph cut based video segmenta- tion system. Recently, graph cut based segmentation tools become prevelant for image/video segmentation problem. However, most of the previous works deal with color information only. Such systems could fail under the condition that there are regions similar in color between foreground and background. Unfortunately, it is usutally hard to avoid. Especially when the objects are filmed under a natural environment. To make it more pratical to use, we propose criterion other than color to conduct the segmentation. Through our observation, motion is a natural choice, since it is usually the case that fore- ground and background has different motion pattern. Moreover, we also ex- tend the Progressive Cut to the temporal-spatial video volume. Experiments shows that by combining color and motion information, our system outper- forms the previous works.