利用像素遮擋關係與圖割技術之深度影像生成

Abstract

在本篇論文中，我們提出了一個創新的立體匹配技術。在我們的方法中，一張參考影像會套用Mean-Sift影像區塊分割技術，而會被分割成若干個自我顏色相似，彼此不互相重疊的區塊。我們套用局部的多視角圖片對應技術和Delaunay三角網格分割技術在特徵點上，用以產生一些在深度空間上的平面，而每一個區塊在深度空間上都能用一個平面來代表。在我們的方法中，多視角圖片對應技術的能量估計函式是以區塊作為單位，而非傳統的以像素為單位。近幾年來，圖像分割技術能夠快速地將一個能量估計函式最佳化，而我們的最佳化策略便是套用建構在以“QPBO”圖像分割為基礎上的alpha-expansion技術。在此我們提出了一個創新的方法，藉由參考影像中像素與其深度的對應關係所建造出的動態深度緩衝區，將區塊的遮擋關係考慮進“QPBO”圖像分割法中。我們拿Middlebury網站上的測試資料去評估我們所提出演算法的準確率，測試結果顯示平均的錯誤像素比例大約在百分之十之內。

In this paper, we propose a new method for stereo matching. In our approach, the reference image is divided into non-overlapping homogeneous segments by Mean-Shift segmentation algorithm. We apply local stereo matching methods and Delaunay triangular on feature points to generate a set of disparity plane for the scene, and each segment is represented as a plane in the disparity space. The stereo matching problem is formulated as a segment-domain energy minimization problem rather than the traditional pixel-domain in our method. The graph cut technique is used to efficiently approximate the optimal solution for energy function in recent years. Our optimization strategy is to apply the disparity plane alpha-expansion, based on “QPBO” algorithm. Therefore we propose an innovative method to approximately encode occlusion consideration into QPBO, based on a dynamic z-buffer constructed by the disparity of the warping pixels in the reference image. We evaluate our algorithm using the cases in the website “Middlebury”, and the results show that the average of bad pixels is about 10%.