分层三维重建学习

CASIA OpenIR > 毕业生 > 硕士学位论文

	分层三维重建学习
	舒茂
	2017-05-10
学位类型	工学硕士
中文摘要	基于图像的三维重建是计算机视觉领域中一个重要的研究主题，分层三维重建是三维重建中的一种重要的方法。分层三维重建首先从图像匹配点得到射影重建，然后由射影重建提升为仿射重建，最后由仿射重建得到度量重建。这种分层次重建的方法简化了重建的复杂度，每一步只需要估计少量的自由参数，从而增强了重建结果的鲁棒性和准确性。近年来深度神经网络和深度学习在物体视觉方面取得了突破性进展，如深度学习在物体检测，物体识别，图像语义分割，图像分类等方面取得了与人类视觉相媲美的性能。但在空间视觉方面，如三维场景重建，视觉物体定位等，深度学习方法的性能却远不如传统的基于几何的方法。究其原因，主要是因为传统基于几何的方法中，可以很方便地集成诸如RANSAC等鲁棒模块从而剔除多幅图像之间不可避免的匹配外点，而基于神经网络和深度学习的途径，目前难以对图像匹配外点进行剔除。针对该问题，本文探究是否能够通过深度学习的方法来剔除图像匹配的外点，从而得到鲁棒的分层三维重建。本文的主要贡献有： 1. 提出了一种学习射影重建的深度卷积神经网络CNN-PR。该网络采用Encoder-Decoder的结构，在无监督学习的框架下，仅仅基于图像特征的对应关系，学习场景的三维射影结构。CNN-PR的输入是一组图像匹配对应点，输出是该组图像点所对应的空间点在射影空间下的齐次坐标表示。CNN-PR集成了外点剔除机制，因此可以从含噪声的对应点中学习得到鲁棒的射影重建。为了验证CNN-PR的性能，本文分别进行了模拟实验和真实实验，并将其与传统的基于几何的方法OpenMVG进行对比。实验结果表明，CNN-PR具有较高的重建精度和很强的鲁棒性。 2. 为了进一步探究学习分层三维重建的可行性，本文提出了一种新的深度卷积神经网络CNN-SR。该网络采用层次化的结构，在监督学习的框架下，逐步从图像对应点中学习得到射影重建、仿射重建、度量重建以及真实的欧氏重建。CNN-SR纳入了降噪机制，可以降低输入中噪声和外点的影响，从而提高三维重建学习的鲁棒性和完整性。为了验证CNN-SR的性能，本文分别进行了模拟实验和真实实验，并与10种传统的基于几何的方法进行了对比分析。实验结果表明，CNN-SR具有较高的重建精度和较强的鲁棒性。
英文摘要	Image-based 3D reconstruction is an important research topic in computer vision，of which stratified 3D reconstruction is an important method. Stratified 3D reconstruction first obtains projective reconstruction from image feature correspondences, then the projective reconstruction is upgraded to affine reconstruction, and finally to metric reconstruction. This layer-by-layer reconstruction method simplifies the complexity of 3D reconstruction, since a small number of free parameters need to be estimated at every step, hence the robustness and accuracy of the reconstruction results are enhanced. In recent years, deep neural networks and deep learning have made tremendous progress in object vision field, for example, deep learning has achieved comparable results to human vision system on object detection, object recognition, image semantic segmentation, image classification etc. However, in spatial vision, such as 3D scene reconstruction, object pose estimation, the learning-based methods are much less successful compared with the currently widely used geometry-based methods. This is because geometry-based methods can conveniently integrate outlier-removal modules, such as RANSAC, which is able to remove matching outliers on the fly, but deep learning based methods lack an embedded outlier-remover. To address this problem, this thesis explores the feasibility of whether deep learning can be used to remove outliers of point correspondences and to achieve robust stratified 3D reconstruction. The main contributions include: 1. We propose a deep convolutional neural network CNN-PR for learning projective reconstruction. This network adopts Encoder-Decoder architecture, and uses unsupervised method to learn the 3D projective structure of scenes from point correspondences across multiple images. The inputs of CNN-PR are a set of putative point correspondences, and the outputs are homogeneous coordinates of corresponding 3D point in projective space. CNN-PR possesses an outlier-removal mechanism so that it can learn robust projective reconstruction from noisy point correspondences. In order to validate the performance of CNN-PR, we conducted simulation experiments and real experiments respectively, and compared CNN-PR with a traditional geometry-based method OpenMVG. Experimental results show that CNN-PR has high reconstruction accuracy and strong robustness. 2. In order to further explore the feasibility of learning stratified 3D reconstruction, a novel deep convolution neural network CNN-SR is proposed. CNN-SR adopts hierarchical network architecture, and uses supervised methods to learn projective reconstruction, affine reconstruction, and metric reconstruction step by step from putative point correspondences across multiple images. CNN-SR possesses a denoising mechanism, which can reduce the influence of outliers and noises, and effectively improve the robustness and completeness of 3D reconstruction results. In order to validate the performance of CNN-SR, we conducted simulation experiments and real experiments respectively, and compared CNN-SR with ten traditional geometry-based methods. Experimental results demonstrate that CNN-SR is both accurate and robust.
关键词	分层三维重建深度学习卷积神经网络外点剔除
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/14685
专题	毕业生_硕士学位论文
作者单位	中国科学院自动化研究所
第一作者单位	中国科学院自动化研究所
推荐引用方式 GB/T 7714	舒茂. 分层三维重建学习[D]. 北京. 中国科学院研究生院,2017.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
template.pdf（165029KB）	学位论文		限制开放	CC BY-NC-SA