Abstract
在本篇論文當中，我們將提出一個新的視覺蛻變方法(View MorphingAlgorithm)，此方法可以產生出一連串視點(View Point)轉換過程中的影像。我們的方法是基於電腦視覺中的仿射空間物件表示法(Affine ObjectRepresentations)，結合電腦圖學中的貼圖方法(Texture Mapping)來產生新的影像。我們的方法除了產生中間過程的影像外，更進一步地也可以容易地加入旋轉、平移等效果。在本篇論文裡，我們做了許多的實驗來說明我們的方法是有用的和有效率。 由於我們建構擬似的模型(pseudo-model)時，不需要知道任的相機參數，所以我們的方法可以應用在任何的照片或是圖畫。加上我們利用電腦圖學的貼圖方法產生影像，不需要計算複雜的光影亮度，而可以得到逼真的影像。所以，我們的方法結合了基於影像的(Image-based)和基於模型的(Model-based)成像優點。我們的方法可以應用在基於影像的虛擬實境上(Image-based VirtualReality)，例如物件導覽系統(Object Movie)或是博物館的簡介系統。並且實作上不需要用到昂貴的電腦硬體，只需要一般的個人電腦就能達到不錯的效果。 因為我們的系統結合影像和模組的優點，所以我們可以將此系統推廣，可以應用在虛擬實境上，讓使用者能與物體有互動關係的建立。將來我們將嘗試將人與物體的互動加入到此系統裡，讓此系統更加地完整。此外，我們將改善本系統的使用者介面，使得我們能夠更便利地操作此系統，讓一般使用者可以方便快速地建立3D的虛擬場景，使得虛擬實境的技術更能普遍推廣到一般日常生活中。In this study, we propose a new view morphing algorithm, whichgenerates a sequence of images to simulate the smooth changes ofview-point. Our algorithm is based on affine objectrepresentations [KOEN 91] [ULLM 91] in computer vision andtexture mapping in computer graphics. Several testingexperiments will be given to show that the proposed algorithmcan generate view morphing sequences effe□□ely andefficiently. Furthermore, the algorithm can easily incorporateobject transformations such as scaling, saring, translation,rotation, etc. Therefore, our algorithm has both merits ofmodel-based and image-based rendering.Recently, there has been a great deal of studies about morphingtechniques. Morphing between two or more images over time is auseful visual technique used in the field of entertainment andvideo games. For example, it has been applied in the productionof the Michel Jackson's MTV, "Black and White", and the movie "Terminator ".The notion of field morphing technique was introduced in [BEIER92]. According to their approach, a set of line segments wereused to specify the corresponding features on two source images.Corresponding pair of line segments on source images determinesa warp from their local coordinate system, u and v, where u isthe position along the line, and v is the distance from theline. When two or more pairs of line segments are specified, theinfluences of line segments are blended by their weightaverages. Theajor drawbacks of this approach are complexity andcontrol. Its complexity is due to that all line segments affectsevery pixel. Moreover, it may generate unexpectedinterpolations, which do not look as natural transitions and noexplicate way of control is given to guarantee the expectedtransitions.To have a better control over the warp, an energy minimizationmethod was proposed to ensure the one-to-one warp among a set offeature points [LEE 95a]. Then an algorithm based on 2D B-splineapproximation was given in [LEE 95b], which is simpler andfaster than the energy minimization method.Although these morphing techniques enable the creation ofeffective image transitions, they do not looks naturally, if wetry to use them to simulate the transitions due to the change ofview point. To cope with these problems Seitz and Dyer [SEITZ96] proposed a new algorithm, called view morphing, whichprewarped two given images, interpolated prewarped images andthen postwarped interpolated images. Their algorithm doesgenerate images in-between with more natural transitions in theaspect of changing vi points. However, it requires the knowledgeof some camera parameters which are hard to obtain outside thelaboratory. In this study, we introduce a new view morphingalgorithm which does not require the knowledge of cameraparameters and can produce quite natural transitions insimulating the change of views.The rest of our thesis is organized as follows. In chapter 2,the basic idea of affine object representations and projectionsare introduced. Then a new view morphing algorithm based onaffine geometry is proposed in chapter 3. The visibility problemof view morphing is studied in chapter 4. First the notions ofpseudo depth and epipolar geometry are introduced to solve theocclusion problem. Then a new morphing technique called v-morphis proposed to solve the unbalanced visibility problem.Experiment resus will be presented in chapter 5 to demonstratethe effect of our approach. The final conclusion and futureworks are given in chapter 6.