以二維影像資訊合成連續的三維景觀

Abstract

本文提供一創新的方法，來模擬三維場景在透視投影下的連續影像變化。藉由少量的已知參考面三維資訊，及至少兩張不同視點的透視投影影像，以透視投影之幾何不變性為基礎，可以產生理論上幾何關係正確的連續三維景觀。與以模型為基礎的影像生成方式比較，本研究省卻大量的三維運算，與三維模型資訊的需求。而較之於以影像為基礎的影像產生方法，因為以幾何不變性為基礎，所以比後者更為精確。本方法假設置放物體的底平面上，至少有五個參考點，可以在這兩張平面影像中定義出來，且任三點不共線。利用視點無關的Cross Ratio值，在底平面上做關於參考點之間的區域定位，以產生物體的影像。本研究將著重在物體幾何框架，隨著視點軌跡移動之影像變化的模擬，並探討影像特徵取樣和特徵定位誤差對成果之影響。

This paper proposes a new method to simulate continuous image transitions of a three dimensional scene under perspective projection. With very limited information about two 3D reference planes and two perspectively projected images from two different viewpoints, continuous image sequences which are geometrically correct can be generated by using some view invariants of perspective projection. This method, compared with model-based algorithms, eliminates most of 3D computations such as coordinate transformations and decreases storage requirements for 3D data of object models. In addition, it has better precision than that in image-based frameworks due to the use of the view invariants. Based on the assumption that there are at least five reference points of the reference plane on which objects are placed can be identified in two corresponding 2D images, and any three of them can not be collinear, images of objects can be synthesized via locating image features on projections of reference plane using the view invariant cross ratios. In this paper, we will focus on the simulation of perspective transition images of object edges due to a linear motion of viewpoint. Error Analysis for the image synthesis due to errors in feature extraction, etc., will also be provided.