基于深度学习的单目图像深度估计

学 学硕士学位论文

Contents

Abstract..................................................................................................I Chapter1Introduction.. (1)

1.1Research background and significance (1)

1.2Current status of research field (2)

1.2.1Depth estimation (2)

1.2.2Deep learning (4)

1.2.3Recurrent Neural Networks (7)

1.3Main research contents of this subject (9)

1.4Overview of manuscript (10)

Chapter2Sequences-based learning depth from a single image (11)

2.1Introduction (11)

2.2Depth estimation using CNN (12)

2.3Convolutional neural network (13)

2.4Superpixel segmentation (17)

2.5Sequences-based depth estimation (18)

2.5.1Decomposition image into set of sequences (20)

2.5.2Strategies for sequences processing (21)

2.6Experiments and results (22)

2.6.1Implementation details (23)

2.6.2Comparisons (24)

2.7Summary (28)

Chapter3Depth estimation using a deep hybrid network (29)

3.1Introduction (29)

3.2Recurrent layer (30)

3.3Long short-term memory (32)

3.4Fully CNN and superpixel pooling (34)

3.5Hybrid network (36)

3.5.1Network architecture (36)

3.5.2Network training and implementation details (38)

3.6Experiments and results (38)

-II-

Contents

3.6.1Comparisions Make3D data (40)

3.6.2Comparisions NUY v2data (44)

3.7Summary (48)

Chapter4Proposed Method Analysis (49)

4.1Analysis of sequences-based approach (49)

4.2Analysis of Hybrid network (49)

4.3Analysis of feature maps (50)

4.4Summary (56)

Conclusions (57)

References (59)

学学位论文 (64)

Acknowledgments (65)

-III-

Chapter1Introduction

Chapter1Introduction

1.1Research background and significance

The science already demonstrated a huge progress in different fields.In the last cen-tury,researchers were looking for a way to build artificial intelligence.They introduced and created methods and algorithms which can be used to model intelligence and ability to learn things.Apparently,building AI is a challenging task which includes a huge amount of various blocks.Each block can be viewed as a different branch of science.Moreover, improvement of the one block is able to make the whole system better.It is easy to see that even small piece of work is important.One interesting research direction that can benefit to artificial intelligence is a scene understanding which is widely used in different fields.

In the modern world,where a relative big of an amount of robots are used in fac-tories and daily life,it is important to teach robots to see as a human do.People receive information about objects around them through the eyes.Moreover,human’s brain is able to process a huge amount information in a short time.In spite of the development of al-gorithms and hardware,modern robots still not able to reach the human level of scene understanding.People not only easily can distinguish objects around them,but also can roughly say how far away specific object from them.Whereas for people,it is a quick task,the same actions can take a relatively big amount of time on modern hardware.

Depth estimation is a significant problem in robotics vision.Different computer vision problems have proven to benefit from the incorporation of depth information.While the big amount of work is focused on depth estimation from stereo images or motion,there has been relative to estimation depth from a single image which is an ill-posed problem, there are an infinite number of world scenes may have produced it.The previous efforts relied on hand-crafted features,such as SIFT,GLOH,HOG,and etc.However,the recent progress in deep learning brings the problem of depth estimation on a new level.

Despite the Microsoft Kinect which has made RGBD images more affordable,the relatively huge amount of datasets is still RGB.Moreover,estimation depth from the sin-gle image can be used for better understanding images distributed over the network and social media,which include many samples of outdoor and indoor scenes;it can be useful for3D reconstruction as well.This had led to wide research interest on the topic of the

-1-