由多張不同視角影像重建物體三維模型

Abstract

我們在這篇論文中，提出了一個從多張視角影像，還原物體三維模型的新方法。此方法還原出的三維模型表面，同時具有”輪廓保留”及” 高顏色一致性”的特質。此方法由兩個階段組成，第一階段使用” 保留輪廓的體積圖切割演算法”，它是由傳統的”體積圖切割方法”修改而成的演算法。雖然它是以體積圖切割方法為基礎，但由於在此演算法中，我們利用了體積圖切割方法的結果回饋調整程式的參數，我們重複著圖形切割與調整參數這兩個步驟，直到產生的結果的輪廓完全符合輸入的影像為止。接著，我們將第一階段得到的物體表面，由第二階段的”最陡坡降演算法”來提高其顏色的一致性。 我們使用了一個真實的物體與兩個合成的物體來測試我 們的方法，改進效果從結果中可以很容易看到，在物體形狀較複雜的狀況下效果尤其顯著。

In this paper, we propose a novel method to recover a 3d object surface with silhouette preserved and high photo consistency properties from multi-view images. Our method is composed of two phases. In the first phase, ”Silhouettes Preserved Volumetric Graph Cuts”, a modified volumetric graph cuts algorithm, is used. This algorithm is on the basis of traditional volumetric graph cuts. However it adjusts the parameters according to the output of volumetric graph cuts and improves the result in an iterative way. These iterative steps will not stop until the obtained 3d surface perfectly matches the observed pictures’ silhouettes. Then, the 3d surface will be refined by Gradient Descent in the second phase. In this phase, the positions of the vertices on the surface will be updated along the normal directions to make the surface’s photo consistency optimized. In the experiment, we test our method with two synthesis models and one real object, and the improvement is obvious, especially in a complicated model case.