以平面為基礎的多視角稠密式三維模型重建

Abstract

本論文透過多台已校正的相機來進行場景或物體的三維稠密式重建，現今有許多方法是運用固定大小的window，從多張影像選兩張影像的window做correlation，來看它們之間的photo consistency。本論文跟這些方法不一樣的是，不再是單純兩兩之間的對應，而是多視角對應，且不使用固定大小的window。重建的步驟是從擷取影像中特徵點開始，方法是用local min或local max的intensity，並使用epipolar line做空間幾何的限制，來得到特徵對應點。在三維空間中，整個物體的表面是由許多平面patch所逼近的，簡而言之，即物體模型是由一堆有向點或點雲所組成的。初始的平面方程式是由特徵對應點或線所計算出來的，接著使用iterative optimization使平面參數的正確性更高，最後完成三維模型。我們運用這重建方法來建合成物來驗證本方法的幾何正確性，並建實拍物來驗證本方法的強韌性。實驗結果比起上面敘述的多視角三維重建方法，看起來更平順，而且更加正確。

This paper addresses 3D dense reconstruction of a scene/object from multiple photos taken by a calibrated camera. Unlike most of the existing methods which use a fixed-sized window to do two-view correlation matching for verifying the photo consistency across the multiple views, our method applies a multi-image matching without requiring a fixed-sized window. The reconstruction begins with extracting image features such as local min/max intensity or the like from the input images, and then applies the epipolar constraints to derive the matched feature pairs. The whole object surface is approximated by a set of planar surface patches in 3D space; that is, the object model consists of a set of oriented points or a point cloud in short. Each planar patch is initially represented by a plane equation derived from the associated line/point feature pairs. Next, an iterative optimization process is called for to refine the plane parameter to attain higher accuracy. Finally, a dense model for 3D model is accomplished. We apply our reconstruction method to synthetic images in order to check the geometry correctness and real images for robustness testing. The experimental results show that our reconstructed models generally look smoother and more accurate than the ones reconstructed by a top existing multi-view stereo method.