杭州电子科技大学：《计算机视觉》课程教学资源（PPT课件讲稿）第五讲 目标分割

第五讲目标分割 周文晖 计算机学院

Hangzhou Dianzi University 杭州电子科技大学 School of Computer Science and Technology 计算机学院 周文晖 第五讲 目标分割 周文晖 计算机学院

课本：目标分割的概念和目的 •概念：将图像划分成若干具有特征一 致性且互不重叠的图像区域的过程。 Aim:to partition an image into a collection of set of pixels >Meaningful regions(coherent objects) >Linear structures (line,curve,... >Shapes(circles,ellipses,... Haug3 hou Dianzi乙niversity抗州电子科技大学 School of Computer Science and Technology计算l学院周文晖

Hangzhou Dianzi University 杭州电子科技大学 School of Computer Science and Technology 计算机学院 周文晖 课本：目标分割的概念和目的 •概念：将图像划分成若干具有特征一 致性且互不重叠的图像区域的过程。 Aim: to partition an image into a collection of set of pixels  Meaningful regions (coherent objects)  Linear structures (line, curve, ...)  Shapes (circles, ellipses, ...)

Other variants:What is segmentation? >Segmentation =partitioning Grouping clustering Carve dense data set into (disjoint)regions Gather sets of items according to some model Divide image based on pixel similarity If items are dense,then essentially the same problem as Divide spatiotemporal volume based on image left.(e.g.,clustering pixels) similarity (shot detection) If items are sparse,then problem has a slightly different Figure/ground separation(background flavor: subtraction) Collect tokens that lie on a line (robust line fitting) Regions can be overlapping (layers) Collect pixels that share the same fundamental matrix (independent 3D rigid motion) Group 3D surface elements that belong to the same surface S.Birchfield,Clemson Univ.,ECE 847,http://www.cos.clomson.odu/-stb/oco847 Hag3 hou Dianzi Universit内抗州电子科技大学 School of Computer Science and Tecfnology计算机学院周文晖

Hangzhou Dianzi University 杭州电子科技大学 School of Computer Science and Technology 计算机学院 周文晖 Other variants：What is segmentation?  Segmentation = partitioning  Carve dense data set into (disjoint) regions  Divide image based on pixel similarity  Divide spatiotemporal volume based on image similarity (shot detection)  Figure / ground separation (background subtraction)  Regions can be overlapping (layers) S. Birchfield, Clemson Univ., ECE 847, http://www.ces.clemson.edu/~stb/ece847  Grouping = clustering  Gather sets of items according to some model  If items are dense, then essentially the same problem as left. (e.g., clustering pixels)  If items are sparse, then problem has a slightly different flavor: • Collect tokens that lie on a line (robust line fitting) • Collect pixels that share the same fundamental matrix (independent 3D rigid motion) • Group 3D surface elements that belong to the same surface

Other variants:What is segmentation? .Segmentation divides an image into groups of pixels .Pixels are grouped because they share some local property(gray level,color,texture,motion,etc.) 0 21 boundaries labels pseudocolors mean colors (different ways of displaying the output) algorithm used:Pedro F.Felzenszwalb and Daniel P.Huttenlocher, Efficient Graph-Based Image Segmentation,IJCV.59(2),2004 Hangzhou Dianzi University抗州电子科技大学 School of Computer Science and Technology计算机学院周文晖 S.Birchfield,Clemson Univ.,ECE 847,http:l/www.ces.clomson.odul-sth/oce847

Hangzhou Dianzi University 杭州电子科技大学 School of Computer Science and Technology 计算机学院 周文晖 Other variants：What is segmentation? •Segmentation divides an image into groups of pixels •Pixels are grouped because they share some local property (gray level, color, texture, motion, etc.) boundaries labels pseudocolors mean colors (different ways of displaying the output) 0 7 3 11 21 3 algorithm used: Pedro F. Felzenszwalb and Daniel P. Huttenlocher, Efficient Graph-Based Image Segmentation, IJCV, 59(2), 2004 S. Birchfield, Clemson Univ., ECE 847, http://www.ces.clemson.edu/~stb/ece847

图像分割的意义 ·区域对于图像理解和识别非常重要，往往表征场景中的目标，或部分目标。 ·一幅图像可以包含多个目标，每个目标包含多个区域，每个区域对应目标的不同部分。 inference Low level features Higher level inference Iaug3 hou Dianzi University杭州电子科技大学 School of Computer Science and Technology计算机学院周文晖

Hangzhou Dianzi University 杭州电子科技大学 School of Computer Science and Technology 计算机学院 周文晖 图像分割的意义 •区域对于图像理解和识别非常重要，往往表征场景中的目标，或部分目标。 •一幅图像可以包含多个目标，每个目标包含多个区域，每个区域对应目标的不同部分。 Low level features Higher level inference inference

图像分割举例 •基于图像亮度 Iaug3 hou Dianzi乙niversit杭州电子科技大学 School of Computer Science and Technology计算l学院周文晖

Hangzhou Dianzi University 杭州电子科技大学 School of Computer Science and Technology 计算机学院 周文晖 图像分割举例 •基于图像亮度

图像分割举例 基于纹理 Iaug3 hou Dianzi University杭州电子科技大学 School of Computer Science and Technology计算l学院周文晖

Hangzhou Dianzi University 杭州电子科技大学 School of Computer Science and Technology 计算机学院 周文晖 图像分割举例 •基于纹理

图像分割举例 光流的概念是gibson在1950年首先提出来的； 是空间运动物体在图像平面上的像素运动的瞬时速度； 基于运动 图像序列中像素在时间域上的变化以及相邻帧之间的相关性来 光流估计 找到上一帧跟当前帧之间存在的对应关系，从而计算出相邻帧 之间物体的运动信息的一种方法。 三维空间的矢量场及其在二维平面内的投影 Hangzhou Dianzi University抗州电子科技大学 School of Computer Science and Technology计算l学院周文晖

Hangzhou Dianzi University 杭州电子科技大学 School of Computer Science and Technology 计算机学院 周文晖 图像分割举例 •基于运动 光流估计 光流的概念是Gibson在1950年首先提出来的； 是空间运动物体在图像平面上的像素运动的瞬时速度； 图像序列中像素在时间域上的变化以及相邻帧之间的相关性来 找到上一帧跟当前帧之间存在的对应关系，从而计算出相邻帧 之间物体的运动信息的一种方法。 三维空间的矢量场及其在二维平面内的投影

图像分割举例 •基于深度 Original image Range image Segmented image (a)Original Image (b)Ground Truth (c)Ours Figure 8.Failure cases on NYUD2 test set Iaug3 hou Dianzi University杭州电子科技大学 School of Computer Science and Technology计算机学院周文晖

Hangzhou Dianzi University 杭州电子科技大学 School of Computer Science and Technology 计算机学院 周文晖 图像分割举例 •基于深度 Original image Range image Segmented image

图像分割应用 Object Category Model Segmentation Cow Image Segmented Cow •理想的图像分割： •能够无指导地、自动分割出完整日标。 Haug3 hou Dianzi乙niversity抗州电子科技大学 School of Computer Science and Technolog)计算机学院周文晖

Hangzhou Dianzi University 杭州电子科技大学 School of Computer Science and Technology 计算机学院 周文晖 图像分割应用 •理想的图像分割： •能够无指导地、自动分割出完整目标。 Segmentation Object Category Model Cow Image Segmented Cow