基于深度学习的数据生成和三维姿态估计

发布时间：2018-01-11 04:28

本文关键词：基于深度学习的数据生成和三维姿态估计　出处：《山东大学》2017年硕士论文　论文类型：学位论文

【摘要】：本文借助深度学习,探索了计算机视觉中一项极其重要而又无比艰难的任务,从单张图片中恢复人体三维姿态。人体姿态是人体建模的基础,很多应用,如姿态识别,动作识别,情感分析等等,都以人体三维姿态估计为基础。同时,人体三维姿态估计又是一项极其困难的任务。在人体三维姿态估计中,需要恢复的信息不止x,y轴,还包括Z轴深度信息。本文提出了通过深度学习,采用卷积神经网络的方法对输入图像进行特征提取,从而恢复出人体三维姿态的方法。卷积神经网络可以最大限度的提取图片的有效特征,在各种计算机视觉的任务上,相比于传统学习方法,表现出了极高的性能提升。然而,在三维姿态估计任务中,人体三维姿态标签特别难标注。这导致了训练集不足,无法训练一个有效的深度神经网络。为解决这个问题,本文提出可以利用计算机图形学中的渲染方法,构造大批量带有三维人体姿态标签的训练数据,从而解决训练集不足的问题。我们发现,在构造渲染数据中,人体姿态空间的覆盖和纹理的多样性是能否成功构造合适训练数据的关键因素。同时,本文提出了一个全自动,大批量的自动渲染算法框架来渲染图片,借助该框架,我们可以从人体姿态空间中进行采样,同时从真实衣服图片中抽取纹理贴图信息。此外,我们还探索了域迁移算法,用于缩小训练数据,也是就是渲染数据,和测试数据,也就是真实数据之间的差异。我们训练了新的网络,针对合成图片和真实图片,可以输出混淆在一起的特征。我们证明,我们的渲染数据可以用于训练卷积神经网络,在人体三维姿态估计上,该网络的表现优于当前最佳方法。同时,域迁移使得合成图片和真实图片的差异进一步缩小,能在原有的基础上进一步提高。
[Abstract]:With the help of deep learning, this paper explores an extremely important and extremely difficult task in computer vision, which is to restore the three-dimensional posture of human body from a single picture, which is the basis of human body modeling and has many applications. For example, attitude recognition, motion recognition, emotion analysis and so on, are based on human three-dimensional attitude estimation. At the same time, human three-dimensional attitude estimation is an extremely difficult task. The information needed to be recovered is not only XY axis, but also Z axis depth information. In this paper, we propose a method of feature extraction from input image by using convolution neural network and depth learning. The convolutional neural network can extract the most effective features of images, compared with the traditional learning methods in various tasks of computer vision. However, in 3D attitude estimation task, it is particularly difficult to label human 3D attitude tag, which leads to the lack of training set. It is not possible to train an effective depth neural network. In order to solve this problem, the rendering method in computer graphics can be used to construct a large number of training data with 3D human pose label. In order to solve the problem of insufficient training set, we find that the cover of human pose space and the diversity of texture are the key factors to construct suitable training data successfully. In this paper, a full-automatic, mass automatic rendering algorithm framework is proposed to render images. With the help of this framework, we can sample from human pose space. At the same time, we extract texture map information from the real clothes image. In addition, we also explore the domain migration algorithm, which is used to reduce the training data, that is, the rendering data, and the test data. That's the difference between real data. We've trained a new network that can produce confusing features for composite and real images. We prove that. Our rendering data can be used to train convolutional neural network, and the performance of the network is better than the current best method in human 3D attitude estimation. At the same time. Domain migration makes the difference between synthetic and real images smaller and can be further improved on the original basis.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP18

【相似文献】