以粒子群最佳化技術重建稠密式物體三維模型

Abstract

本論文旨在利用Particle Swan Optimization的方式重建Multi-view Stereo多視角三維模型。本論文主要延伸patch based的expansion重建方式，分為幾大部分，分別為camera calibration、seed patch optimization、patch expansion、patch filtering，並且透過PSO來提供一個derivative free的解法。 本論文特色之一是延伸patch based的概念並改進在較sparse的camera view之下提供完整重建，主要使用geometric與texture matching兩個限制來定義較多且可靠的visible camera，並利用GLN-PSO local search的特性來進行有效率patch optimization與expansion。此外對於textureless的物體，我們使用multi-scale的pyramid image，相較於直接擴增patch size能夠提供更快的收斂速度。而對於非自然物的重建，我們則是利用adaptive fitness weighting來進行patch optimization，以獲得預期的銳利轉折處。

This paper presents a stochastic optimization based multi-view stereo (MVS) approach for 3D dense reconstruction. We propose to apply adaptive weighted stereo matching functions to achieve more accurate optimization result. On the other hand, the reconstruction completeness falls short of the lack of enough visible views. We advocate allowing the child patch to borrow the parent visible view when needed even though the parent view is not in the specified viewing angle range. In addition, we shall adopt a GLN-PSO stochastic patch optimization method to avoid the local traps of a derivative based numerical optimization method. To improve the reconstruction quality we propose a patch priority queue to select the best patch to search for the next patch for the patch expansion process.