在異質多核心系統上的混合式深度圖優化方式

Abstract

三維（3D）的內容在當前的大眾媒體越來越普遍，藉由增加深度資訊到傳統媒體內容中，其立體效果能讓觀眾感覺物體像是現實般的生動。這種全新的時尚視覺刺激大大提高3D內容的製作和成像技術的需求。 3D影像的處理現今大多是以複數的鏡頭做最初的畫面擷取，以極幾何(epi-polar geometry)[1]的概念經由深度估算(disparity estimation)演算法[2]來計算出各個位置的深度。產生出來的深度資訊我們稱為深度圖，可以利用這些資訊增加傳統影像的訊息量。在3D的成像技術中像是3D顯示器便是使用這些深度圖的資訊來合成並呈現出立體效果。 在生成3D的資訊時，深度圖的估計和優化步驟在深度估算中是最耗時的部分。全域最小化的方法可以得到更好的質量，但同時有計算時間長的問題。區域最小化的方法有巨大的計算平行性，但可能會導致質量低劣。本論文提出了一種自我調適的混合式深度圖優化方式，同時從全域和區域的方法得到好處，並減輕相互的缺點。我們所提出的混合方法在異質的多核心系統上執行可以達到平均12.78倍加速並只有4.3％的質量下降。

Three dimensional (3D) contents are becoming prevalent in the current mass media. By augmenting the depth information to the conventional media contents, the stereoscopic effects make viewers feel like watching vivid objects in real life. The brand new fashion of vision stimulus is considerably raising the demands on the production and presentation technologies for 3D contents. Generating images with depth information is an enabling technology to create 3D experiences. One of the most widely used methods today is based on a multi-view scheme. The target scene is captured by several cameras where each camera is located at a different view point. The depth information of the target scene is then calculated by combining the image information from multiple view points and calculating the geometry [1] with a depth estimation algorithm [2]. The generated depth information, which is usually referred as a depth map, is augmented to the conventional image. With the depth map, the 3D presentation technology, such as 3D displays, will then use this information to synthesize and present the contents with stereoscopic effects. The optimization of the depth map is the most time consuming part of the depth estimation process when generating 3D contents. The global minimization approach could return better quality while suffering from long computation time. The local minimization approach has huge computation parallelism, but could result in inferior quality. This thesis proposes a self-adjusted hybrid scheme which benefits from both global and local approaches while alleviating the corresponding drawbacks. By executing on a heterogeneous many-core system, the proposed hybrid approach can demonstrate an average of 12.78x runtime speedup with only 4.3% quality degradation.