本篇論文提出一個基於superpixel的2D-to-3D的方法，此為自動深度擷取轉換方法。近來三維立體影像需求的增加，而三維影像內容資源之缺乏。如果想容易的享受到逼真的立體視覺效果，勢必需要開發低成本、高效率的轉換方法，將原本二維的影像快速的轉換成三維立體影像。 首先，我們使用高斯模型作前景偵測，分離出前景與背景，接著，我們使用superpixel演算法來找出邊緣資訊，我們將顏色相近和位置相鄰的pixels作clustering，根據superpixel群聚出來的像素我們給予初始的深度值，我們會初始六種不同的深度圖，利用hough transform來找出消失線的斜率，接著利用斜率可知哪個深度圖是我們要的，給完初始深度值後，我們再用sobel edge detection來作第二次的邊緣偵測，用兩種不同的閥值來得到不同邊緣資訊，一個有較多雜訊但邊緣資訊較完整，另一個雜訊較少但邊緣資訊也較缺乏，然後用thinning演算法來降低邊緣像素的寬度使其變成只有1 pixel，比較這兩個結果後重新賦予深度值，接著再將前景資訊加進來給前景物件相同的深度值，為了使深度圖更加精準，因此，我們利用四種方向掃描整張影像來修正深度值，即可得到最後的深度圖，最後，再用depth image based rendering (DIBR)來合成左右視角的影像，如此，就完成了3D影像。 ;This paper proposes novel method for 2D-to-3D video conversion. It is based on boundary information to automatically generate the depth map. First, we use Gaussian model to detect foreground objects and then separate the foreground and background. Next, we use the superpixel algorithm to find the edge information. Then according to the pixels which are clustered by superpixel, the initial depth values are acquired. Based on the result for depth value assignment, we detect the edges by Sobel edge detection with two thresholds to strength the edge information. To identify the pixel of boundary, we use thinning algorithm to the results of edge detection. Comparing these results and re-assign the depth value, the depth value of foreground will be refined. In order to make more accurate depth map, we use four kinds of scanning path for the entire image to correct depth values. After that, we will have the final depth map. Finally, use depth image based rendering (DIBR) to synthesize left and right view image. The 2D-to-3D conversion will complete. Combining the depth map and the original 2D video, a vivid 3D video is produced.
