使用高解析度資料集於改善天空影像去背方法

Abstract

這篇論文提出了一個改善的天空影像去背方法。我們提出了一個高解析度的天空影像去背資料集，透過人工標註三分圖 (trimap) 並導入基於三分圖的影像去背網路，生成精確的天空透明遮罩 (alpha matte)。接著，我們修改了 PP-Matting，一種通用的影像去背網路，使其在天空影像去背上有更好的表現。具體來說，我們將 CoordConv 層和可訓練的導向濾波器整合到網路中，並進行初步研究以找到最佳設計。實驗結果展示，相較於現有的天空影像去背資料集，我們的資料集為各種影像去背網路提供了更好的訓練環境，無論是在量化指標還是視覺品質上。此外，我們修改過的網路架構在天空影像去背上，相較於現有方法具有更優秀的表現。

This paper proposes an improved method for sky image matting. We present a high-resolution sky image matting dataset and generate accurate alpha mattes by leveraging a trimap-based image matting network with manually annotated trimaps. Next, we modify the architecture of PP-Matting, a general-purpose image matting network, to better suit the task of sky image matting. Specifically, we integrate CoordConv layers and a trainable guided filter into the network and conduct a preliminary study to find the optimal design. Experimental results demonstrate that our dataset provides a better training environment for various image matting networks compared to an existing sky image matting dataset in both quantitative metrics and visual quality. Additionally, our modified network shows superior performance compared to existing methods.