Skeleton-driven Animation Transfer

隨著電影、網路遊戲、模擬系統的發達，三維立體模型現今已經成為我們生活中不可或缺的一部份。而要使一個三維立體模型能夠作出動作，最常被使用的方法就是必需要由美工人員先替此模型建出一套骨幹，並將模型上的所有網格與此骨幹建立出影響權重關係，之後在每一個關鍵影格調整骨幹的姿勢，使三維立體模型作出一連串的動畫。這樣的過程是很繁瑣的，此外，一般在網際網路上我們可以很容易地就搜尋到釵h三維立體模型，但是這些三維模型幾乎都是只具有網格的結構，而沒有附上可用來編輯動作的骨幹，因此就算我們想要將一模一樣的動作套到不同的三維立體模型上，還是得重覆之前所述的工作。 在此論文中，我們提出了利用參數化找出目標及來源兩個三維立體模型的關係，並根據此關係來作動作複製，使得原本僅具有網格結構的目標三維立體模型，也能夠產生與來源三維立體模型一樣架構的骨幹及權重關係，並作出類似的動作。整個演算法分成幾大部份，首先必需半自動選出兩個三維立體模型相對應的重要特徵，之後再將三維立體模型作平面的參數化，並依照先前所找出的重要特徵作對齊，之後根據此參數化平面，可以找到兩個三維立體模型所有網格的對應關係。一旦找到了這項對應，就可以重建出一樣架構的骨幹，並將所有與動畫相關的資訊都轉移過去，以完成動作的複製。 這套系統旨在幫助美工人員能夠很快速地先將相同的動作套用到各種不同的三維立體模型上，並產生一樣架構的骨幹及權重關係，之後，我們的系統能夠將新產生出來的動作轉成Maya的檔案格式，以供美工人員到Maya裡面去微調每個關鍵網格。此外，由於這套系統的介面及操作流程相當簡單，因此針對像對使用Maya這樣複雜軟體不是很熟悉的初學者而言，是相當方便直覺的。As the fundamental building blocks of movies, networking computer games, and simulation systems, 3D models have become an important part of our daily life recently. The common solution for driving a 3D model move is building a skeleton and setting the binding weight for it by artist, and then key-pose editing is necessary. Even when we want to create the same motion sequence for different 3D models, we still have to repeat the process. So it is very lengthy and time consuming to create an imposing scene such as full of soldiers or dancers. Hence we propose an efficient system to clone a skeleton-driven model to another new character in this paper. To clone the motion of a source model, which motion is driven by its skeleton, to another one, the basic idea is that we have to construct the same skeleton structure for the target model. Since the motion data is always saved as the attributes of the vertices, in order to transfer the skeleton and all of the motion data to the target model, we utilize surface parameterization to find the correspondence of all vertices between source and target models. Using this correspondence, we also can transfer the information including color, texture, skeleton and binding weight from the source model to the target one. The whole process was abstracted as follows. First, the user needs to dissect both of the two models into homologous patches and manually mark some common features of them. Since the patches dissected from the two models are disk-like, they can be mapped to a plane using discrete conformal mapping. In order to compute better correspondences between a pair of homologous patches, we use a foldover-free warping scheme to align those features. Then the correspondence of all vertices can be obtained by overlaying the two aligned embeddings. To transfer a skeleton, we select two vertices as the markers for each joint of the skeleton and remember their numerical relation to the joint. Because of the consistent correspondence, we can reconstruct the skeleton of the same structure for target model. Once the motion data is transferred, our system can clone the motion from the source model to the target one. Our system aimed at speeding up and facilitating the process of cloning animation to different 3D models. We can efficiently transfer skeleton and binding weight by only mark few vertices manually. Our system can convert transferred motion data to the Maya file format, so that the artists can refine the cloned key poses in Maya.Chapter 1 Introduction 1 1.1 Motivation 1 1.2 System Overview 3 1.3 Previous Work 4 1.3.1. Decompose into Patches 4 1.3.2. Planar Parameterization 9 1.3.3. Spherical Parameterization 11 1.3.4. Feature alignment 11 1.4 Animation Data 12 1.4.1. Skeleton 12 1.4.2. Binding Weight 12 1.4.3. Key-Frame Poses 13 1.5 Contribution 14 1.6 Organization 15 Chapter 2 System Pipeline 17 2.1 System Input/Output 17 2.2 ASF/AMC/WGT File Format 17 2.2.1. ASF 17 2.2.2. AMC 18 2.2.3. WGT 19 2.3 System Work Flow 19 2.3.1. Previous Prepare 19 2.3.2. Kernel 20 2.3.3. Refinement 20 Chapter 3 Features Suggestion 21 3.1 Introduction 21 3.2 Find Tip Features by ReebGraph 21 3.3 Find Join Features by Decomposition 25 3.4 Refine Manually 27 Chapter 4 Parameterization 29 4.1 Introduction 29 4.1.1. Coarse-to-fine spherical parameterization and warping over sphere 29 4.1.2. Spring spherical parameterization 31 4.2 Conformal map 33 4.3 Combine Relaxation to Free Fold-over 34 4.4 Feature Alignment 35 4.4.1. Use Radial Basis Function 37 4.4.2. Use Fold-over Free Warping Scheme 37 4.5 Correspondence Representation 39 4.6 Results 40 Chapter 5 Skeleton Transfer 47 5.1 Introduction 47 5.2 Reconstruct Skeleton 47 5.3 Results 49 Chapter 6 Animation Clone 51 6.1 Introduction 51 6.2 Animation Clone 51 6.3 Results 52 Chapter 7 Performance and Results 55 Chapter 8 Conclusion and Future Works 61 8.1 Conclusion 61 8.2 Future Works 61 References 6