High Quality Novel View Generation of Object Movies

雖然真實物體的數位化已經發展了好一段時間，至今仍然只有少數為了產生高品質畫面而被提出來的方法，例如博物館通常需要高品質的數位文物來展示古文物，然後，目前並沒有任何方法有考慮到了實際上在相機校正及立體建模時是會出現誤差的。我們使用了隨視點改變的景深圖來克服這些誤差，同時結合隨視點改變的景深圖及隨視點改變的材質貼圖，我們可以產生出比以往更加好的結果，我們產生出來的影像品質幾乎可以達到原始照片的品質。們利用我們先前所提出的方法中得到的立體模型來有效的去得到一組為每個視點合身製作的景深圖，因此，即便是精細的物體，也可以被忠實的記錄下來，也可以被正確的描繪出來，在此同時，結合了隨視點改變之材質貼圖及我們所建立的隨視點改變的景深圖，我們就可以生成出近乎原始影像品質的新視點圖，我們可以藉此建立起既有影像間所沒被拍攝到的影像，而得到一個能連續轉動的環物影片。由於我們產生的影像品質幾乎與原始影像差不多，所以我們的方法非常適用於像是展示博物館文物這種要求高品質影片的應用。Modeling of real objects has been developed for a long time. However, there are still few methods which are designed for applications requiring extremely high quality results, such as applications in museum. None of existing methods consider errors in camera calibration and modeling. We use a view-dependent depth map to compensate for these errors. At the same time, combining our depth maps and view-dependent texturing technique generates better quality rendering results than ever before. The quality of our rendering results are close to original photographs.ur framework utilizes the global polygon model to obtain fitted view-dependent depth maps efficiently. At the same time, combining our depth maps and view-dependent texturing technique gives accurate and photo realistic results. Our method carries out the seamless transition between native photos and novel rendered images. Therefore, our work is especially suitable to virtual exhibiting of the museum artifacts.1 Introduction 7 Related Work 12 Depth Map Re nement Framework 16.1 System Overview 16.2 Matching Cost 17.3 Depth Re nement Algorithm 20.4 Re nement I : Re nement with Automatically Marked Pixels 25.5 Re nement II : Re nement with Manually Speci ed Constraints 28.5.1 Two Kinds of Constraints 30.5.2 Depth Map Propagation 34.6 Re nement III : Local Re nement 37 Rendering 41.1 Disparity-based View Morphing 41.2 Point Splatting 42.3 Fusion of Views 45 Experiments and Results 49 Conclusion and Future Work 60.1 Conclusion 60.2 Future Work 62 Texture Space Matching 64eferences 6