臺灣博碩士論文加值系統

本論文提出一個利用自應性門檻來找出可信賴的區域平面，再以此些平面參數代入區域平面掃描，進而在相當短的時間內得到圖上幾個最佳的深度值。我們先用Harris角點和Daisy描述子來標記有足夠特徵的點，接下來用我們提出的自應性門檻來分組這些點群。我們還加了貪婪演算法以及以面為初始出發點的分群法則，這使得我們的演算法可以在幾個迭代內就達到收斂的結果。非自應性門檻的優點在於我們可以讓每個點群集自己去找出適合本身的點，而不受限傳統固定門檻的限制。最後套用了次全域的匹配方法，來找出有遮蔽的地方，最後可以對每個點提出次像素的精準度，使得最後結果極具參考價值。

We propose an algorithm that can provide disparity candidates of each pixel in the given input images in stereo matching problems. Most of the stereo matching algorithm can be separated in 2 parts: initial planes generation and optimization. We add some ideas trying to improve the former, and especially, we put emphasis on local plane generation. Our method estimates the planes from sparse feature matches, and we propose a strategy in collecting adjacent features with adaptive error threshold for plane fitting. The parameters of each plane is re-estimated whenever new feature points are incorporated according to the adaptive threshold setting. Furthermore, we apply the greedy aggregation: the plane with more features from the previous iteration is more probable for assigned features. We extract main planes fitting the scene structure without the burden of tuning the plane fitting error threshold. Finally we will apply semi-global matching to obtain the top five accurate candidates for each pixel.

致謝 i
中文摘要 ii
ABSTRACT iii
LIST OF FIGURES iv
LIST OF TABLES vii
CONTENTS viii
Chapter 1 Introduction 1
1.1 Thesis Organization 2
1.2 Contributions 2
Chapter 2 Related Work 4
2.1 Daisy Dense Descriptors 5
2.1.1 Introduction of Daisy Dense Descriptors 5
2.1.2 The parameters of Daisy Dense Descriptors 7
2.2 Local Plane Estimation 8
2.2.1 Greedy Algorithm 9
2.2.2 Adaptive Threshold 10
2.3 Semi Global Matching 11
Chapter 3 System Overview 14
3.1 Pre-Processing 15
3.2 Sparse Features Extraction 16
Chapter 4 Local Plane Generation 17
4.1 Initialization and The First Iteration 18
4.2 Iterative Assignment 20
4.3 Halting Criteria 24
Chapter 5 Local Plane Sweeps 27
5.1 Matching Cost 28
5.2 Smoothness Cost 29
5.3 Semi Global Matching 30
Chapter 6 Result and Conclusion 33
6.1 Top-N Accuracy 34
6.2 Results 35
6.3 Conclusion 37
BIBLIOGRAPHY 38

[1] D. Scharstein, R. Szeliski, and R. Zabih, “A taxonomy and evaluation of dense two-frame stereo correspondence algorithms.” in Proceedings of the IEEE Workshop on Stereo and Multi-Baseline Vision (SMBV’01), SMBV ’01, pages 131–, Washington, DC, USA, 2001.

[2] S. Sinha, D. Scharstein and R. Szeliski, “Efficient High-Resolution Stereo Matching using Local Plane Sweeps” in Proceeding CVPR ''14 Proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition Pages 1582-1589

[3] V. Kolmogorov and R. Zabih. “Computing visual correspondence with occlusions using graph cuts.” in International Conference for Computer Vision, pages 508-515, 2001.

[4] Y. Boykov, O. Veksler, and R. Zabih. Ef cient approximate energy minimization via graph cuts. in IEEE Transactions on Pattern Analysis and Machine Intelligence, 23(11):1222-1239, 2001.

[5] E. Tola, V. Lepetit, and P. Fua. A fast local descriptor for dense matching. in CVPR, 2008

[6] M. Brown, G. Hua, and S. Winder. Discriminative learning of local image descriptors. In TPAMI, 33(1):43–57, 2011.

[7] H. Hirschmuller. Stereo processing by semi-global matching and mutual information. In TPAMI, 30(2):328–341, 2008.
[8] D. G. Lowe. Distinctive image features from scale-invariant keypoints. In IJCV, 60(2):91–110, 2004.