基於隨機最佳化之高精確度及完整度多視角影像密集式三維模型重建

Abstract

本論文提出了基於隨機最佳化演算法的多視角影像之密集式三維模型重建方法，三維模型重建的關鍵在於如何提昇高精確度及完整度，本論文針對此關鍵問題提出解決方法。首先，重建過程會因在多視角影像間的立體匹配計算錯誤造成精確度下降。大部份的立體匹配計算錯誤發生於低紋理、深度不連續處、或是重覆性紋理區域，本論文提出了適應性權重立體匹配函式來處理前述問題。另一方面也會容易因缺乏足夠的視角影像，造成重建完整性不足，尤其是在多視角影像並非規律擺放時，部份區域僅落於少數視角影像中。針對此議題，本論文令漸進式擴張重建時發生此問題的平面延用其擴張前的平面資訊，讓只有少數視角影像所觀看到的區域能有較多的資訊來進行空間定位重建。 另外，數值微分最佳化方法常因限制條件而難以進行微分，而且也常面臨落入區域最小值的問題，例如在使用共軛梯度法時，所使用的初始解不佳時，容易無法得到最佳解。在本論文中，我們使用GLN-PSO隨機式最佳化方法解決前述問題，此外也易於針對較複雜的適應性權重立體匹配函式進行最佳化。為了改善重建的品質，本論文提出透過優先隊列來選擇最佳的擴展重建過程。另一方面，我們基於多視角輪廓線影像提出了一個可經由參數調整物體模型重建品質的方法，該方法有著高效率及節省計算所需記憶體的特性。透過擷取輪廓線重建的物體表面，可以有效地改善原本的密集式三維模型重建結果的準確度及完整度。 上述提出的方法皆經過合成影像及真實影像進行實驗比較，真實影像亦包含Middlebury標準資料集。從實驗結果可展示出目前的實作結果，與其他演算方法進行比較後具有相當的競爭力與後續發展潛力。

This dissertation addresses a stochastic optimization approach for multi-view 3D dense reconstruction. In pursuit of better accuracy and completeness of a dense 3D reconstruction model we provide the valuable techniques to tackle the key problems unsolved. First, the accuracy downgrading is attributed to the stereo miss-matches across multiple views. Most of stereo matching errors occur in the regions of less texture, depth discontinuity, or repeated patterns. We propose to apply adaptive support weight stereo matching functions to deal with these problems. On the other hand, the reconstruction completeness falls short of expectation due to the lack of enough visible views. In particular, in the case of a non-uniform camera deploy, often insufficient camera views are available in certain specified viewing range. We advocate allowing the child patch to borrow the parent view when needed, even though the parent view is not in the specified viewing 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 such as the conjugated gradient method with a poor initial solution. To improve the reconstruction efficiency and quality we propose a patch priority queue to provide the best patch for the patch expansion process. Furthermore, we also present a novel octree reconstruction method to extract a set of exact surface feature points based on multi-view silhouette images, which can be used to enhance the patch based reconstruction. The proposed method is tested on synthetic and real image data sets. The real data sets include the dinosaur and temple data sets in the well-known Middlebury benchmarks and a high-resolution face data set. The experimental results demonstrate the current implementation results are encouraging in comparison with the top ranked reconstruction methods reported in the public Middlebury MVS website.