基於單張影像之城市景觀三維深度估測技術研究

Abstract

在本篇論文中，我們著重於重建單張影像的深度。我們考慮由城市場景裡取得線性透視資訊藉以瞭解場景的3-D模型。不同於使用物體之間遮蔽關係來估計影像相對深度但忽略透視幾何的方法，我們的目標是結合透視幾何資訊及遮蔽關係的資訊，透過垂直和水平方向深度梯度地圖的建構以提供深度的變化趨勢。在我們的做法中，首先將影像切割成區塊，並分類成主要的幾何類別：垂直物，地面和天空。藉由提取影像中的消失點，我們利用消失點和區塊間相對位置以及區塊類別來產生初始的深度梯度地圖。接著我們再由主要方向的消失線和遮蔽邊界的資訊，利用條件隨機場模型來調整初始深度梯度地圖；而修正的深度梯度地圖即為求出的最佳解。最後由結合兩方向深度梯度地圖產生圖片的深度。在城市場景中，我們的方法可估測出不錯的深度圖。

In this thesis, we focus on recovering a depth map from a single image. Given an image of urban scenes, we extract linear perspective information to establish the 3-D scene model. Unlike these approaches which use occlusion relationship between objects to estimate the relative depth of the image, we further combine the perspective geometry information with the occlusion relationship between objects. In our approach, we construct depth gradient maps for both vertical and horizontal directions to represent the depth variation trend for the image. To accomplish this, the image is first partitioned into components, which are classified into three geometric classes: vertical plane, ground plane, and sky. By extracting the vanishing point of the image content, we generate initial depth gradient maps based on the relative position between the vanishing point and the classified components. After that, we use the main directions of vanishing lines and non-occlusion boundaries to revise the initial depth gradient maps by using a CRF (conditional random field) model. A refined solution for depth gradient maps is generated by finding the optimal solution. The depth map can be generated by integrating depth gradient maps. We demonstrate that this approach can produce pleasant depth maps for images of urban scenes.