计算机视觉基础学习课件.ppt

单击此处编辑母版标题样式,单击此处编辑母版文本样式,第二级,第三级,第四级,第五级,#,计算机视觉基础,1,目录,概述,成像模型,图像滤波,边缘检测,特征检测与识别,光流,迹线几何与立体视觉,Structure from Motion,大数据驱动的视觉计算,2,2025/12/16 周二,1.,概述,3,2025/12/16 周二,相关研究领域,计算机图形学：模型,图像,4,2025/12/16 周二,相关研究领域,计算摄影学：图像,图像,5,2025/12/16 周二,计算机视觉,图像,模型,让计算机“看懂”图像和视频,这是何种场景,?,汽车在哪里,?,建筑物有多远,?,6,2025/12/16 周二,视觉是自然智能不可思议的技艺,猕猴的大脑皮层中视觉部分占据大约,50%,人脑中有关视觉的部分所占比重最大,这是皇后还是象,?,视觉,7,2025/12/16 周二,安全,健康,监控,家务,进入,娱乐,计算机视觉为什么重要？,8,2025/12/16 周二,1966:Minsky,给本科生布置了一个计算机视觉的暑假作业,1960s:,合成的虚拟世界的理解,1970s:,图像理解方面的进步,1980s:,几何和精度,1990s:,人脸识别,;,统计分析开始流行,2000s:,更多的识别,;,大规模标记数据集可用,;,开始视频处理,Guzman 68,Ohta Kanade 78,Turk and Pentland 91,计算机视觉简史,9,2025/12/16 周二,计算机视觉的应用：,OCR,数字识别,AT&T,实验室,周二,计算机视觉的应用：人脸检测,目前许多数码相机都能检测人脸,Canon,Sony,Fuji,11,2025/12/16 周二,计算机视觉的应用：笑脸检测,Sony Cyber-shot T70 Digital Still Camera,12,2025/12/16 周二,计算机视觉的应用：由成千上万的图像重建三维,13,2025/12/16 周二,计算机视觉的应用：物体识别,(,超市中,),LaneHawk by EvolutionRobotics,“A smart camera is flush-mounted in the checkout lane,continuously watching for items.When an item is detected and recognized,the cashier verifies the quantity of items that were found under the basket,and continues to close the transaction.The item can remain under the basket,and with LaneHawk,you are assured to get paid for it“,14,2025/12/16 周二,计算机视觉的应用：基于视觉的生物测量,12,岁,30,岁,15,2025/12/16 周二,计算机视觉的应用：无密码登录,笔记本电脑和其他设备上的指纹扫描仪,人脸识别系统,16,2025/12/16 周二,计算机视觉的应用：物体识别,(,手机上,),17,2025/12/16 周二,黑客帝国,计算机视觉的应用：特效,-,形状捕获,18,2025/12/16 周二,加勒比海盗,计算机视觉的应用：特效,-,运动捕获,19,2025/12/16 周二,计算机视觉的应用：体育,Sportvision,first down line,Nice,explanation,on,周二,计算机视觉的应用：智能汽车,Mobileye,汽车上的视觉系统，如,BMW,、,GM,、,Volvo,等,21,2025/12/16 周二,计算机视觉的应用：,Google,汽车,22,2025/12/16 周二,计算机视觉的应用：太空视觉,视觉系统的几项任务：,全景图缝合,三维地形建模,障碍检测，位置跟踪,其他,(,参阅,Matthies,等人的,“,Computer Vision on Mars,”),NASA,的火星探索计划：,2007,年精神号漫游车,23,2025/12/16 周二,计算机视觉的应用：工业机器人,视觉引导的机器人给汽车上定位螺母,24,2025/12/16 周二,计算机视觉的应用：机器人,机器人足球赛,NASA,的火星漫游车,斯坦福生活机器人（洗碗）,25,2025/12/16 周二,计算机视觉的应用：医学成像,手术导航,3D,核磁共振、,CT,26,2025/12/16 周二,2.,成像模型,27,2025/12/16 周二,计算机视觉的相关研究领域,计算机图形学：模型,-,图像,计算摄影学：图像,-,图像,计算机视觉：图像,-,模型,28,2025/12/16 周二,图像形成,设计一个相机：,思路,1:,将底片放在物体前方,我们能得到一幅合适的照片吗？,29,2025/12/16 周二,针孔相机,思路,2:,增加一个障碍物阻止大多数的光线,减少模糊,光圈控制光线量,30,2025/12/16 周二,针孔相机,f,f=,焦距,c=,相机中心,c,31,2025/12/16 周二,暗箱,:,相机前身,中国（公元前,470,年）和希腊（公元前,390,年）,暗箱,UNC Chapel Hill,的暗室,Photo by Seth Ilys,32,2025/12/16 周二,第一张照片,现存的最老照片,花了,8,小时在锡盘上成像,Joseph Niepce,1826,第一张照片的照片,保存在,UT Austin,33,2025/12/16 周二,维度降低的机器（,3D,到,2D,）,3D,世界,2D,图像,34,2025/12/16 周二,投影的欺骗性,35,2025/12/16 周二,射影几何,丢失了什么,?,长度,哪个球更近些,?,谁更高,?,36,2025/12/16 周二,长度没有被保留,B,C,A,37,2025/12/16 周二,射影几何,丢失了什么,?,长度,角度,垂直,?,平行,?,38,2025/12/16 周二,射影几何,什么被保留,?,直线依然是直线,39,2025/12/16 周二,消逝点和消逝线,物理世界中的平行线在图像中相交于“消逝点”,40,2025/12/16 周二,消逝点和消逝线,o,消逝点,1,o,消逝点,2,消逝线,41,2025/12/16 周二,消逝点和消逝线,消逝点,消逝线,消逝点,垂直消逝点,(,无穷远处,),42,2025/12/16 周二,消逝点和消逝线,43,2025/12/16 周二,投影：世界坐标,图像坐标,Camera Center(t,x,t,y,t,z,),.,.,.,f,Z,Y,.,Optical Center(,u,0,v,0,),v,u,44,2025/12/16 周二,齐次坐标,变换,齐次图像坐标,齐次场景坐标,由齐次坐标转换回来：,转换到齐次坐标：,45,2025/12/16 周二,齐次坐标,齐次坐标是缩放不变量,笛卡尔坐标中的点在齐次坐标中是一条射线,齐次坐标,笛卡尔坐标,46,2025/12/16 周二,齐次坐标的基本几何学,直线方程,:ax+by+c=0,给像素坐标增加分量,1,得到齐次坐标,两点叉积得到一条直线,两条直线的叉积得到这两条直线的交点,47,2025/12/16 周二,齐次坐标解决的另一个问题,笛卡尔坐标,:,(Inf,Inf),齐次坐标,:,(1,1,0),平行线求交,笛卡尔坐标,:,(Inf,Inf),齐次坐标,:,(1,2,0),48,2025/12/16 周二,投影矩阵（针孔相机模型）,x,:,图像坐标,(u,v,1),K,:,内部矩阵,(3x3),R,:,旋转矩阵,(3x3),t,:,平移量,(3x1),X,:,世界坐标,(X,Y,Z,1),O,w,i,w,k,w,j,w,R,T,49,2025/12/16 周二,投影矩阵,K,内部假设：,单位宽高比,光心坐标,(0,0),无倾斜,外部假设：,无旋转,相机坐标,(0,0,0),50,2025/12/16 周二,移除“已知光心”的假设,内部假设：,单位宽高比,无倾斜,外部假设：,无旋转,相机坐标,(0,0,0),51,2025/12/16 周二,移除“正方形像素”假设,内部假设：,无倾斜,外部假设：,无旋转,相机坐标,(0,0,0),52,2025/12/16 周二,移除“无倾斜”的假设,内部假设：,外部假设：,无旋转,相机坐标,(0,0,0),53,2025/12/16 周二,允许相机移动,内部假设：,外部假设：,无旋转,54,2025/12/16 周二,点的三维旋转,围绕坐标轴的,逆时针,旋转：,p,p,g,y,z,55,2025/12/16 周二,允许相机旋转,56,2025/12/16 周二,自由度,5,6,57,2025/12/16 周二,消逝点,=,无穷远点的投影,58,2025/12/16 周二,正射投影,透视投影的特例,正交投影的中心到图像平面的距离为无穷大,也称作,“,平行投影,”,其投影矩阵是什么,?,Image,World,59,2025/12/16 周二,比例缩放的正射投影,透视投影的特例,物体面积相对于到相机的距离来说很小,也称为,“,弱透视,”,其投影矩阵是什么,?,Image,World,Slide by Steve Seitz,60,2025/12/16 周二,视场,(,缩放,),61,2025/12/16 周二,假设有两个三维的立方盒子放在地上，面朝观察者，一个近，一个远,透视图中它们看起来是什么样子？,在弱透视中它们看起来又是什么样子？,62,2025/12/16 周二,针孔相机之外,:,径向失真,桶形失真校正,无失真,桶形失真,枕形失真,63,2025/12/16 周二,3.,图像滤波,64,2025/12/16 周二,图像滤波,空间域图像滤波,直接对像素进行操作,平滑化、锐化,频率域图像滤波,修改图像的频率,去噪、采样、图像压缩,模板和图像金字塔,将模板匹配到图像,检测、粗糙到精细,65,2025/12/16 周二,Image filtering,图像滤波：计算每个位置处局部邻域的函数值,滤波很重要,!,图像增强,去噪、调整大小、对比度增强，等等,从图像中提取信息,纹理、边缘、特征点，等等,检测模式,模板匹配,66,2025/12/16 周二,例：箱式滤波器,1,1,1,1,1,1,1,1,1,67,2025/12/16 周二,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,0,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,0,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,68,2025/12/16 周二,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,0,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,0,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,69,2025/12/16 周二,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,0,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,10,20,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,0,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,70,2025/12/16 周二,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,0,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,10,20,30,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,0,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,71,2025/12/16 周二,0,10,20,30,30,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,0,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,72,2025/12/16 周二,0,10,20,30,30,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,0,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,?,73,2025/12/16 周二,0,10,20,30,30,50,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,0,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,?,74,2025/12/16 周二,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,90,0,90,90,90,0,0,0,0,0,90,90,90,90,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,90,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,10,20,30,30,30,20,10,0,20,40,60,60,60,40,20,0,30,60,90,90,90,60,30,0,30,50,80,80,90,60,30,0,30,50,80,80,90,60,30,0,20,30,50,50,60,40,20,10,20,30,30,30,30,20,10,10,10,10,0,0,0,0,0,1,1,1,1,1,1,1,1,1,75,2025/12/16 周二,箱式滤波器,每个像素的值用其邻域像素的平均值替换,实现平滑效果（去除尖锐特征）,1,1,1,1,1,1,1,1,1,76,2025/12/16 周二,箱式滤波的平滑效果,77,2025/12/16 周二,线性滤波器,0,0,0,0,1,0,0,0,0,原始图像,?,78,2025/12/16 周二,线性滤波器,0,0,0,0,1,0,0,0,0,原始图像,滤波结果,（无变化）,79,2025/12/16 周二,线性滤波器,0,0,0,1,0,0,0,0,0,原始图像,?,80,2025/12/16 周二,线性滤波器,0,0,0,1,0,0,0,0,0,原始图像,往左移动,1,个像素,81,2025/12/16 周二,线性滤波器,原始图像,1,1,1,1,1,1,1,1,1,0,0,0,0,2,0,0,0,0,-,?,（注意：滤波器各元素之和为,1,）,82,2025/12/16 周二,线性滤波器,原始图像,1,1,1,1,1,1,1,1,1,0,0,0,0,2,0,0,0,0,-,锐化滤波器：,突出与局部均值之差别,83,2025/12/16 周二,锐化,处理前,处理后,84,2025/12/16 周二,其他滤波器,-1,0,1,-2,0,2,-1,0,1,垂直边缘,（绝对值）,Sobel,算子,85,2025/12/16 周二,其他滤波器,-1,-2,-1,0,0,0,1,2,1,水平边缘,（绝对值）,Sobel,算子,86,2025/12/16 周二,我们改怎样合成运动模糊？,theta=30;len=20;,fil=imrotate(ones(1,len),theta,bilinear);,fil=fil/sum(fil(:);,figure(2),imshow(imfilter(im,fil);,87,2025/12/16 周二,高斯滤波器,邻域像素根据其接近程度计算贡献值,0.003 0.013 0.022 0.013 0.003,0.013 0.059 0.097 0.059 0.013,0.022 0.097 0.159 0.097 0.022,0.013 0.059 0.097 0.059 0.013,0.003 0.013 0.022 0.013 0.003,5 x 5,=1,88,2025/12/16 周二,高斯滤波器的平滑效果,89,2025/12/16 周二,高斯滤波器,移除图像中的高频分量（低通滤波）,图像变得更光滑,自卷积是高斯滤波的一种形式,因此，可用小宽度内核进行平滑，如此重复，得到用大宽度内核相同的结果,用宽度为,的高斯内核卷积两次，等同于用宽度为,2,的高斯内核卷积一次,可分离的核,二维高斯可分解成两个一维高斯的积,90,2025/12/16 周二,高斯滤波器的可分离性,二维高斯可表示为两个函数的乘积，一个是,x,的函数，另一个是,y,的函数,这两个函数都是一维高斯,91,2025/12/16 周二,可分离性的例子,*,*,=,=,二维卷积,(,仅中心位置,),滤波器分解为两个,一维滤波器的乘积,沿行执行卷积计算,紧跟着沿列进行卷积,92,2025/12/16 周二,例,:,混合图像,高斯滤波器,拉普拉斯滤波器,高斯函数,单位脉冲,高斯函数的拉普拉斯,93,2025/12/16 周二,4.,边缘检测,94,2025/12/16 周二,边缘检测,目标,:,识别图像中的突变（不连续）,直观上，图像的大多数语义和形状信息都可由边缘信息,表示,比像素更紧凑,理想,:,画家的线条画（对象级）,95,2025/12/16 周二,为什么我们要关注边缘,?,提取信息、识别目标,恢复几何和视点,Vanishing,point,Vanishing,line,Vanishing,point,Vertical vanishing,point,(at infinity),96,2025/12/16 周二,图像边缘的来源,边缘是由各种因素引起的,深度不连续,曲面颜色不连续,光照不连续,曲面法向不连续,97,2025/12/16 周二,边缘特写,Source:D.Hoiem,98,2025/12/16 周二,边缘的特点,一条边是图像亮度函数中突变的地方,图像,亮度函数,（沿水平扫描线）,一阶导,边缘对于导数的极值,99,2025/12/16 周二,亮度轮廓线,100,2025/12/16 周二,带有少量高斯噪声,梯度,101,2025/12/16 周二,噪声的影响,考虑图像中单行,/,单列,哪里是边缘,?,102,2025/12/16 周二,解决方案,:,先平滑,为了检测到边缘，查找 的峰值,f,g,f*g,103,2025/12/16 周二,计算如下,:,卷积微分定理,f,104,2025/12/16 周二,平滑可去除噪声，但会模糊边缘,在不同尺度下检测到边缘,1,个像素,3,个像素,7,个像素,平滑和局部化之间的权衡,105,2025/12/16 周二,观察：沿着“脊”处的梯度幅值较大,我们该怎样检测到实际的边缘点？,我们怎样将这些边缘点连接起来形成边缘曲线？,算法实现,106,2025/12/16 周二,边缘检测器的设计,一个好的边缘检测器遵循的原则：,有好的检测率,最优的检测器应该能检测出所有真实边缘，而忽略噪声和其他瑕疵,有好的局部化效果,检测到的边缘必须尽可能的靠近真实边缘,对于每个真实边缘点，检测器必须返回一个点,边缘检测线索,颜色、亮度、纹理的变化,连续性,高层知识（语义等）,107,2025/12/16 周二,Canny,边缘算子,计算机视觉中用得最广的边缘检测器,理论模型：加性噪声干扰的阶梯边缘（,Step-Edge,）,Canny,展示了高斯函数的一阶导紧密逼近最优化信噪比和局部化的边缘算子,J.Canny.,A Computational Approach To Edge Detection,.IEEE Transactions on Pattern Analysis and Machine Intelligence,8:679-714,1986.,108,2025/12/16 周二,例,原始图像（,Lena,）,109,2025/12/16 周二,高斯滤波器的导数,X,方向,Y,方向,110,2025/12/16 周二,计算梯度,(DoG),X,方向导数,Y,方向导数,梯度幅值,111,2025/12/16 周二,每个像素的梯度方向,theta=atan2(gy,gx),112,2025/12/16 周二,对每个方向进行非最大值压缩,若点,q,的值大于其梯度方向上的点,p,和点,r,的值，则认为,q,处为极大值,点,p,和点,r,的值通过插值得到,113,2025/12/16 周二,非最大值压缩之前,114,2025/12/16 周二,非最大值压缩之后,115,2025/12/16 周二,滞后阈值化,检测梯度值的最大值是否足够大,在边缘曲线起始处使用大阈值,后续使用小阈值,116,2025/12/16 周二,Canny,边缘检测结果,117,2025/12/16 周二,Canny,边缘检测算子,高斯函数的,x,和,y,方向导数滤波图像,计算梯度的幅值和方向,非最大值压缩：,细化多个像素宽的“脊线”,阈值化和连接：,定义高,/,低两个阈值,使用高阈值开始一条边缘曲线，低阈值用于后续,118,2025/12/16 周二,的影响,(,高斯核的宽度,),原始图像,值的选择依赖于需求：,大的,值用于检测大尺度边缘,小的,值用于检测细节特征,119,2025/12/16 周二,5.,特征检测与识别,120,2025/12/16 周二,特定识别任务,121,2025/12/16 周二,场景分类,户外,/,室内,城市,/,森林,/,工厂,/,等等,122,2025/12/16 周二,图像标注,街道,人,建筑,山,旅游业,多云,砖,123,2025/12/16 周二,目标检测,检测行人,124,2025/12/16 周二,图像解析,mountain,building,tree,banner,market,people,street lamp,sky,building,Svetlana Lazebnik,125,2025/12/16 周二,场景理解,?,126,2025/12/16 周二,可变性,:,相机位置,光照,形状参数,类内变化,?,识别就是建模可变性,127,2025/12/16 周二,类内变化,128,2025/12/16 周二,识别研究的历史,1960s early 1990s:,几何时代,Svetlana Lazebnik,129,2025/12/16 周二,可变性,:,相机位置,光照,q,对齐,Roberts(1965),;Lowe(1987);Faugeras Huttenlocher&Ullman(1987),形状：假设已知,130,2025/12/16 周二,对齐,对齐：在两幅图像的特征对之间，通过一个变换来进行拟合,找到一个变换,T,，使得下式最小化,T,x,i,x,i,131,2025/12/16 周二,识别,成为一个对齐问题：组块世界,J.Mundy,Object Recognition in the Geometric Era:a Retrospective,2006,L.G.Roberts,Machine Perception of Three Dimensional Solids,Ph.D.thesis,MIT Department of Electrical Engineering,1963.,132,2025/12/16 周二,对齐：,Huttenlocher&Ullman(1987),133,2025/12/16 周二,表示和识别物体类别是更难的,.,ACRONYM(Brooks and Binford,1981),Binford(1971),Nevatia&Binford(1972),Marr&Nishihara(1978),134,2025/12/16 周二,通过部件进行识别,Primitives(geons),Objects,Biederman(1987),135,2025/12/16 周二,Zisserman et al.(1995),Generalized cylinders,Ponce et al.(1989),Forsyth(2000),通用形状基元,?,136,2025/12/16 周二,识别研究的历史,1960s early 1990s:,几何时代,1990s:,基于表观的模型,137,2025/12/16 周二,图像可变性的经验模型,基于表观的技术,Turk etc.,138,2025/12/16 周二,Eigenfaces(Turk&Pentland,1991),139,2025/12/16 周二,颜色直方图,Swain and Ballard,Color Indexing,IJCV 1991.,140,2025/12/16 周二,全局表观模型的局限性,要求模式全局注册,对于背景混乱、有遮挡、以及几何变换不鲁棒,141,2025/12/16 周二,识别研究的历史,1960s early 1990s:,几何时代,1990s:,基于表观的模型,Mid-1990s,:,滑动窗口方法,142,2025/12/16 周二,滑动窗口方法,143,2025/12/16 周二,识别研究的历史,1960s early 1990s:,几何时代,1990s:,基于表观的模型,Mid-1990s:,滑动窗口方法,Late 1990s:,局部特征,144,2025/12/16 周二,物体识别的局部特征,D.Lowe(1999,2004),145,2025/12/16 周二,大规模图像搜索,结合局部特征、索引和空间约束,146,2025/12/16 周二,大规模图像搜索,结合局部特征、索引和空间约束,Philbin et al.07,147,2025/12/16 周二,大规模图像搜索,结合局部特征、索引和空间约束,148,2025/12/16 周二,识别研究的历史,1960s early 1990s:,几何时代,1990s:,基于表观的模型,Mid-1990s:,滑动窗口方法,Late 1990s:,局部特征,Early 2000s:,零件,-,形状模型,149,2025/12/16 周二,模型,:,物体视作零件集合,零件间的相对位置,零件的表观特征,零件,-,形状模型,150,2025/12/16 周二,星座模型,Weber,Welling&Perona(2000),Fergus,Perona&Zisserman(2003),151,2025/12/16 周二,Representing people,152,2025/12/16 周二,识别研究的历史,1960s early 1990s:,几何时代,1990s:,基于表观的模型,Mid-1990s:,滑动窗口方法,Late 1990s:,局部特征,Early 2000s:,零件,-,形状模型,Mid-2000s:bags of features,153,2025/12/16 周二,Object,Bag of words,Bag-of-features,模型,154,2025/12/16 周二,识别研究的历史,1960s early 1990s:,几何时代,1990s:,基于表观的模型,Mid-1990s:,滑动窗口方法,Late 1990s:,局部特征,Early 2000s:,零件,-,形状模型,Mid-2000s:bags of features,当前趋势,:,局部和全局相结合的方法、数据驱动的方法、上下文方法,155,2025/12/16 周二,数据驱动方法,J.Hays and A.Efros,Scene Completion using Millions of Photographs,SIGGRAPH 2007,156,2025/12/16 周二,数据驱动方法,J.Tighe and S.Lazebnik,ECCV 2010,157,2025/12/16 周二,D.Hoiem,A.Efros,and M.Herbert.,Putting Objects in Perspective,.CVPR 2006.,几何上下文,158,2025/12/16 周二,判别训练的基于零件的模型,P.Felzenszwalb,R.Girshick,D.McAllester,D.Ramanan,Object Detection with Discriminatively Trained Part-Based Models,PAMI 2009.,159,2025/12/16 周二,6.,光流,160,2025/12/16 周二,光流的概念,光流：图像序列（时间域和空间域）亮度变化的矢量场函数。,161,2025/12/16 周二,运动和感知组织,有时候，运动是唯一线索,162,2025/12/16 周二,运动和感知组织,甚至“贫瘠的”运动数据也能引发很强的感知,G.Johansson,“,Visual Perception of Biological Motion and a Model For Its Analysis,Perception and Psychophysics 14,201-211,1973.,163,2025/12/16 周二,运动的用途,估算三维结构,基于运动线索分割对象,学习和跟踪动态模型,识别事件和活动,改进视频质量（,motion stabilization,）,164,2025/12/16 周二,运动场,运动场是三维场景运动到图像的投影,非旋转球朝相机移动，其运动场是什么样子的？,165,2025/12/16 周二,光流,定义：光流是图像中亮度模式的,表观运动,理想地，光流应该与运动场一致,必须注意：,表观运动可能由照明改变而非实际运动引起,例如：一个在固定照明条件下均匀旋转的球面,vs.,一个在移动光照条件下的静止球面,166,2025/12/16 周二,Lucas-Kanade,光流算法,基本算法,多分辨率算法,迭代算法,167,2025/12/16 周二,image I,image J,图像,1(t),的高斯金字塔,图像,2(t+1),的高斯金字塔,image 2,image 1,从粗到细光流估算,run iterative L-K,run iterative L-K,warp&upsample,.,.,.,168,2025/12/16 周二,例,Khurram Hassan-Shafique,CAP5415 Computer Vision 2003,169,2025/12/16 周二,多分辨注册,170,2025/12/16 周二,光流计算结果,171,2025/12/16 周二,光流计算结果,*From Khurram Hassan-Shafique,CAP5415 Computer Vision 2003,172,2025/12/16 周二,7.,极线几何与立体视觉,173,2025/12/16 周二,多视角,Hartley and Zisserman,Lowe,多视角几何、匹配、不变特征、立体视觉,174,2025/12/16 周二,为什么要多视角,?,单视角中，结构和深度是模棱两可的,175,2025/12/16 周二,为什么要多视角,?,单视角中，结构和深度是模棱两可的,光心,P1,P2,P1=P2,176,2025/12/16 周二,什么线索可以帮助我们感知三维形状和深度？,177,2025/12