基于深度学习的单目图像深度估计

来源 :西南交通大学 | 被引量 : 0次 | 上传用户:wang1hnsc
下载到本地 , 更方便阅读
声明 : 本文档内容版权归属内容提供方 , 如果您对本文有版权争议 , 可与客服联系进行内容授权或下架
论文部分内容阅读
One of the main goal of computer vision is the image understanding. Despite the recent success in different tasks, such as object recognition, pose estimation, etc., some of them still remain formidable, however. Much progress has been made in extracting primitives from image and understanding it. Nevertheless, modern robotic systems still have not reached the human level of scene understanding, because algorithms have a high complexity and require a huge amount of modern hardware.Depth estimation is a significant task in the robotics vision. A wide range of vision problems has proven to benefit from the incorporation of the depth information. In this thesis we address the depth estimation from a single monocular image, which is an illposed problem because there are an infinite number of world scenes may have produced the given image. When a huge amount of research was focused on depth estimation from a stereo images or motion, depth prediction from a single image has attracted attention in recent years.To address our main objective, we develop several deep convolutional neural networks that can be applied to predict depth from the image. The first network predicts the depth of a sequence of superpixels. The main idea is process image into the sequential manner, which can decrease network training time. The second hybrid network can be applied to predict the depth of the whole image. The main difference of the current networks for depth prediction is that our network is composed of convolution and recurrent layers, which are composed of the Long Short-Term Memory units(LSTM). The LSTM unit is famous for the ability to memorize long-range context, which can be applied to feature maps and lead to better feature predictions.Our experiments on benchmark dataset demonstrated, that hybrid network can be efficiently applied to predict a depth of an image. Moreover, the composition of recurrent and convolutional layers provide more satisfied results. In addition, we investigate the difference between a standard and hybrid network.
其他文献
随着嵌入式技术在工业控制领域的应用日益普及,对于嵌入式管理软件的研究也越来越成为一个热门的课题。因此,如何利用现有的硬件平台,研究嵌入式系统管理软件的开发流程,对相
细分方法是曲线曲面几何造型中的常用方法,它可以克服参数表示方法的局限性,并具有从离散到离散、规则简单、易于修改和极限曲线曲面良好的光滑性等优点,广泛应用于三维几何
非负矩阵分解(Non-negative Matrix Factorization,NMF)是一种较新的矩阵分解方法,它将给定的一个非负矩阵分解为左右两个非负矩阵因子的乘积,可得到被分解矩阵的低秩逼近。
三维表面建模技术是科学计算可视化领域中的热点研究问题,在地质勘探、医学诊断等方面都有着重要的应用。因此,对三维表面建模技术的研究,具有重要的学术意义和应用价值。本
随着互联网的发展,P2P(Peer-to-Peer)对等网已成为目前研究的热点。P2P网络是利用终端闲散资源,包括计算能力、存储空间、网络带宽等提供服务,从而实现快速资源搜索,海量数据
兼有基础模式无线局域网(WLAN,Wireless Local Area Networks)单跳传输和移动Ad Hoc网络(MANET,Mobile Ad Hoc Network)多跳路由优点的无线Mesh网络(WMN,Wireless MeshNetwork)是一种
当今,世界经济从工业化转向信息化,从工业社会到信息社会的转变已经成为不可逆转的大趋势。在这个过程中,经济发展越来越需要信息的支持。工作流技术应势而起,它使得人们关注
校园一卡通系统已经在很多高校被广泛使用,在校园卡的日常应用过程中会产生很庞大的数据量,这些信息包含着校园卡中许多不为人知的特点。校园卡的使用过程中,存储了学生每一
近几年来,嵌入式技术和多媒体通信技术迅速发展。嵌入式IP视频会议系统凭借低成本和高质量的优势成为研究的焦点。凭借简单、可扩展和易实现等特点,SIP逐步成为3G和NGN领域首
近几年来,随着网络的普及,对等网络(Peer-to-Peer, P2P)技术飞速发展。P2P系统由于方便、快捷和开销少的优点受到了大多数网络用户的欢迎。文件共享系统是P2P技术最为广泛的