室內模型最佳視角選擇系統

Abstract

室內場景照片是一種幫助使用者快速了解室內格局的一種媒介，常被用於租屋、室內設計、廣告等商業應用上，符合審美觀的室內照片會吸引讀者目光進而影響商業上的產值。過去對於3D 模型選擇最佳視角的研究主要聚焦在單一模型上，他們利用 curvature 等區域性的資訊尋找 3D saliency map，以最 saliency 的位置為中心拍攝，但對於一個包含多個模型的室內場景不只要考慮模型之間的關係，還會受到室內空間的限制影響拍攝的角度與範圍。為了克服這些問題，我們先找出室內空間中 saliency 的部分視為拍照主體，在人類視覺系統對於對比度較敏感的前提下，計算顏色出現頻率及在 L*a*b*色彩空間中的差距，並利用空曠程度定義結構上的 saliency 地圖，就可以得到具有顏色及結構的 saliency 位置。在選擇最佳視角的演算法中，參考室內攝影學技巧，並計算離散化後的拍照視角是否符合畫面的審美法則，最後將視角分數算出來後經過排名提供給使用者參考。

Interior photograph is the most efficient way to help people realize indoor scenes. It is commonly use in house renting, indoor scene design, advertisement and several commercial areas. Clearly, attractive photos would enhance the value of products. To achieve the aim, previous methods estimate 3D saliency maps with localinformation such as curvatures to determine good views. However, these methods often focus on a single object and the relations between furnitures are not considred if the goal is to take good photographs of an indoor scene.In contrast, the presented system not only estimates 3D saliencies of a scene but also considers several photography principles when determining the best view. Specifcally, textures, geometries, and spatial relations are taken into account to identify salient objects. After that, a score function that consists of blocking, free space, and rule of third are maximized to select the best view of an indoor scene. Experimental results demonstrate the effectivenss of our technique.