MPEG視點合成參考軟體 於NVIDIA CUDA之加速與改進

Abstract

隨著立體影像技術的盛行，自由視點視訊(FTV, Free Viewpoint Television)成為熱門的研究主題。任意視點合成為一關鍵技術，其中視點合成的即時輸出和複雜度降低為重要技術議題。由於NVIDIA 公司提出之Compute Unified Device Architecture (CUDA)平台能有效的處理資料密集型的應用程式，因此我們將MPEG提供的視點合成參考軟體 (View Synthesis Reference Software, VSRS)[2]移植至此架構上。為了在CUDA平台上實現VSRS，必須加強VSRS的平行度。在實作過程中，我們兼顧平行度與影像品質。首先，我們提出內部空洞填補(Intra Hole Filling)方法，取代原本中值濾波器的方法。此外，我們為了減少資料的相依性，透過CUDA 執行緒對資料做適宜的分割，並且重新安排資料執行順序，以減少分支指令。綜合上述技巧，我們節省超過94％的計算時間，並且得到類似的影像品質。

With the prosperity of 3D technology, Free Viewpoint Television (FTV) becomes a popular research topic. “View Synthesis” is a key step in FTV. There are some important and to-be-solved issues such as real-time operation and complexity reduction. NVIDIA Compute Unified Device Architecture (CUDA) is an effective platform in handling data-intensive applications. To implement the MPEG view synthesis reference software (VSRS) on CUDA, we parallelize the VSRS structure. In the meanwhile, our proposed parallel scheme improves the picture quality. We first propose an intra hole filling scheme to replace the original median filter. Then, to avoid data dependence we properly partition the data so that they can be processed by the parallel GPU threads. Also, we rearrange the data processing order in the threads to reduce branching instructions. Combining these techniques together, we save more than 94% computing time and achieve a similar image quality.