基於區域可變式視窗大小和適應性權重的視差估計演算法

Abstract

立體視差估計演算法被廣泛的利用在許多實際應用層面，像是3D視訊會議和多視角立體電視等。在一個典型的地區式視差估計演算法當中，通常會使用一個固定的視窗大小來聚合視差匹配的代價。然而，越大的視窗大小雖然能提升在低紋理區塊的表現，卻也同時讓視差的邊界被模糊化。 在本篇論文中，我們提出了一個可變動的視窗大小，先利用顏色資訊把圖片連結或切割成多個區塊，再利用區塊的資訊來決定視窗的大小，同時在最後的優化步驟，這些資訊也能用在一個十字區域式表決機制。另外，為了讓計算結果的品質更加提升，我們更結合了微型普查(mini-census)和顏色差距的比對方式。從實驗結果顯示，我們的方法能夠有效提升原本方法的效能，而且經由Middlebury網站的評估，我們的方法是目前的地區式演算法當中名列前茅的。

Stereo matching algorithm has been widely adopted by various stereo vision applications such as 3D video conference and free viewpoint TV. In the typical local methods for stereo matching, a fixed support window size is often adopted in cost aggregation step. Larger supporting window can improve the stereo matching performance at low texture regions. However, it is blurred near depth discontinuities. In this paper, we propose a variable window size selection strategy before cost aggregation step. The strategy determines the support window by utilizing the segment information derived from a color based segmentation method. This information is also used for region-based cross voting scheme in refinement step. Moreover, a combined matching cost measure with mini-census and color difference is proposed. Experimental results show that the proposed method effectively improves the performance of original method. According the performance evaluation at the Middlebury website, the proposed method is one of the current state-of-the-art local methods.