基於深度神經網路之圖層三維深度排序技術

Abstract

在這篇論文裡，我們提出一套方法用來找出單張影像的三維深度。不同於利用場景幾何資訊來推論深度的做法，我們不對場景有任何的幾何假設，且只擷取影像裡小區域的深度特徵來幫助我們找到深度。在我們的做法中，首先將影像切割成數個區塊，我們主要專注在小區域裡的深度排序。利用深度神經網路去分類小區域裡深度排序的類別，且自動學習出適合的特徵，然而利用傳統的方式學習深度神經網路會面臨梯度消失的問題。因此，我們的學習過程有兩部分：無監督式預先學習以及有監督式微調模型。學習完深度神經網路後，對於一張影像我們把影像切割成小區域進行測試。然而小區域對於深度排序的結果會不一致。因此，我們結合小區域的相對深度資訊並且把這些資訊轉換成有向圖。再藉由找出最小反饋邊集合以得到整體一致的深度排序結果。

In this thesis, we propose a method to estimate 3-D depth map from a single image. Unlike approaches employing a geometric model behind the scene to infer the depth map, we propose to estimate the 3-D scenes by extracting local depth cues without any structure assumptions in the scene instead. In our approach, first we partition the image into several regions. We focus on inferring the depth ordering in a local patch. Then we apply the model of deep neural network to figure out which depth order class the local patch belong and to automatically learn the appropriate features in our problem. However, training the deep neural network by classical method will encounter the vanishing gradient problem. To tackle the problem, our training algorithm consists two phase: the unsupervised pre-training phase and the supervised fine-tuning phase. After training the deep neural network, we test the image by feeding the local patch into the deep neural network. In practice, the resultant depth orders that are sorted by the deep neural network and from different local patches may be contradictory. Hence, we combine these local depth order reasoning to construct a direct graph. By finding the minimum feedback arc set, we can obtain a depth order with global consistency.