没有合适的资源?快使用搜索试试~ 我知道了~
首页轻量级宽基线街景图像插值:单应传播与优化
轻量级宽基线街景图像插值:单应传播与优化
0 下载量 20 浏览量
更新于2024-08-26
收藏 2.04MB PDF 举报
"这篇研究论文探讨了宽基线街道图像插值中的单应传播与优化技术,旨在解决在不依赖复杂3D重建或深度网络的情况下,实现高效且精确的图像插值问题。" 在计算机视觉领域,宽基线街景图像插值是一个重要的课题,它涉及到如何将两个或多张具有大视差的街景图像融合成一个连续的全景图像。然而,由于视差大带来的挑战,如物体变形、重叠和遮挡,使得这一任务极具难度。现有的方法往往依赖于重量级的3D重建算法或计算量庞大的深度学习网络,这在实时性和计算效率上存在限制。 论文的作者Nie, Zhang, Sun, Su和Li提出了一种轻量级且高效的解决方案,该方案利用简单的单应性计算和精炼操作来估计输入视图之间分段平滑的单应性变换。单应性是一种几何关系,可以描述平面场景中二维图像间的对应关系。他们通过可靠和不可靠的超像素鉴别,将单应性拟合与单应性传播相结合,提高了估计单应性变换的准确性和鲁棒性。这种方法的优势在于,它可以区分哪些区域的单应性变换更可靠,从而更有效地处理图像间的几何转换。 随后,论文引入了单应性约束的网格变形概念,提出了一种新颖的单应性约束扭曲形式化方法。该方法利用扭曲网格的一阶连续性来强制相邻单应性之间的平滑过渡,减少了重叠、拉伸和压缩等小瑕疵,从而提高插值结果的质量。 通过这种方法,论文提出的算法不仅提高了插值的准确性,还减少了计算复杂度,使得在不需要大规模3D重建或深度学习网络的情况下,也能实现高质量的宽基线街景图像插值。这对于实时应用,如自动驾驶、无人机航拍图像处理以及虚拟现实等有着显著的实际意义。
资源详情
资源推荐
Wide-Baseline View Synthesis. Wide-baseline view synthesis
can be achieved if assuming enough additional information
or constraints. For example, Google StreetView [11] required
depth information from laser scanners. Wide-baseline stereo
methods [10], [26], [27] relied on binocular images. Recent
methods of image-based rendering [12], [13], [28] recon-
structed 3D model of a scene, and compensated for the errors
of the reconstruction by silhouette-aware warping [12] or
depth synthesis [13], etc. However, they demand a collection
of images of the scene as input, and are not robust enough for
rural scenes (such as those of KITTI [2]) where 3D reconstruc-
tion may fail. Recently, some researchers [1], [29], [30] tried to
apply learning methods to novel view synthesis. For exam-
ple, Flynn et al. [1] developed a DeepStereo system that
trained a deep network to predict novel views from the
world’s imagery. The learning-based method required a
large amount of training data and much training time. Cam-
era poses were also required when training. By contrast, our
method only needs two images as input, which is more light-
weight and flexible.
3APPROACH
Our aim is to smoothly transform from a source street view
I
s
to a corresponding target view I
t
. Our method divides I
s
into superpixels, and views each superpixel as a small plane
which corresponds to somewhere in the target image I
t
by a
homography. We try to compute the homographies of
superpixels as accurate and robust as possible. As shown in
Fig. 1, we propose a three-step homography computing and
refining algorithm: (1) reliable and unreliable superpixel
discrimination: a superpixel is identified as either reliable or
unreliable; (2) homography fitting and propagating: the
homographies of reliable superpixels are computed by
homography fitting while the homographies of unreliable
superpixels are computed by homography propagation; (3)
homography optimization: the homographies are further
optimized to enforce smoothness between adjacent homog-
raphies. We compute both the forward homographies from
I
s
and I
t
and the backward homographies from I
t
to I
s
.
With the estimated homographies, intermediate images can
be interpolated and blended.
3.1 Preparation
Before introducing our three-step algorithm, we do some
preparations in this section, as shown in the second column
of Fig. 1. We oversegment the source image I
s
into a collec-
tion of superpixels. We also compute the Approximate
Nearest Neighbor Field from I
s
to I
t
which helps to discrim-
inate reliable superpixels from unreliable ones.
We employ the method of [31] for superpixel segmenta-
tion. Superpixel serve as the building block of the proposed
algorithm. Since a superpixel usually has homogeneous
color, we can view each superpixel as a small plane and
compute a homography for it. In this paper, we set the
superpixel size at 500 pixels. For a 800600 image, there are
about 1,000 superpixels. We denote the set of superpixels of
an image by S¼fS
i
ji 2f0; ...;n 1gg.
We use the generalized PatchMatch method [32] to com-
pute the ANNF from I
s
to I
t
which is a dense correspon-
dence field indicating for each patch of I
s
the most similar
patch in I
t
. A source patch can be translated, scaled, and
rotated to find the best match in the target image. The image
at the bottom of the second column of Fig. 1 illustrates the
computed ANNF.
3.2 Reliable and Unreliable Superpixel
Discrimination
Given the superpixels and ANNF computed in the last
section, we identify whether a superpixel is reliable or not
for homography fitting by measuring the consistency of the
ANNF of the superpixel. The ANNF indicates pixel corre-
spondences between the source and target images. If the
correspondences of the pixels of the whole superpixel are
consistent enough with each other, we view the superpixel
as reliable.
We employ and improve the method of [33] to measure
the reliability of a superpixel. First, we define two pixels as
consistent if their ANNF correspondence vectors are similar.
Specifically, let P and Q be two patches of image I
s
, and P
0
and Q
0
be the corresponding patches of image I
t
by ANNF,
with p, q, p
0
, and q
0
be the centers of the four patches and
p
0
¼ p þFðpÞ, q
0
¼ q þFðqÞ, where F indicates the ANNF.
Normally, to compute the consistency error Cðp; qÞ between
Fig. 1. Overview of our algorithm that computes for a source image piecewise smooth homographies with respect to a target image. Given a pair of
images with wide baseline, we first segment the source image into superpixels and compute the Approximate Nearest Neighbor Field (ANNF) from
the source image to the target image in the preparation step. We then discriminate reliable superpixels from unreliable superpixels by the consistency
of ANNF. Next, we compute homographies for reliable superpixels by homography fitting, and compute homographies for unreliable superpixels by
homography propagation. Finally, the estimated homographies undergo an optimization to enforce smoothness between them.
2330 IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 23, NO. 10, OCTOBER 2017
剩余13页未读,继续阅读
weixin_38675506
- 粉丝: 4
- 资源: 931
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
最新资源
- 彩虹rain bow point鼠标指针压缩包使用指南
- C#开发的C++作业自动批改系统
- Java实战项目:城市公交查询系统及部署教程
- 深入掌握Spring Boot基础技巧与实践
- 基于SSM+Mysql的校园通讯录信息管理系统毕业设计源码
- 精选简历模板分享:简约大气,适用于应届生与在校生
- 个性化Windows桌面:自制图标大全指南
- 51单片机超声波测距项目源码解析
- 掌握SpringBoot实战:深度学习笔记解析
- 掌握Java基础语法的关键知识点
- SSM+mysql邮件管理系统毕业设计源码免费下载
- wkhtmltox下载困难?找到正确的安装包攻略
- Python全栈开发项目资源包 - 功能复刻与开发支持
- 即时消息分发系统架构设计:以tio为基础
- 基于SSM框架和MySQL的在线书城项目源码
- 认知OFDM技术在802.11标准中的项目实践
资源上传下载、课程学习等过程中有任何疑问或建议,欢迎提出宝贵意见哦~我们会及时处理!
点击此处反馈
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功