本课程面向国内外所有从事计算机视觉、模式识别、图像处理、以及相关基础学科(应用数学、统计、计算机)领域的本科毕业班学生、研究生、博士后、及青年教工。讲座为期两周,分成几个主题，系统化地介绍计算机视觉和模式识别的基础和前沿理论、原理、方法、最新进展。

课程安排

时间： 2006年7月3-16日(1-2日报到)

地点：湖北鄂州, 莲花山风景区，莲花山计算机视觉与信息科学研究院

主讲： 朱松纯教授，美国洛杉矶加州大学, 统计系与计算机系

其它讲员：吕宏静博士，美国洛山矶加州大学，统计系与心理系博士后(2006)；香港大学，心理系助理教授 (2007)；

吴郢教授, 美国西北大学，电子工程与计算机科学系；

屠卓文博士, 美国洛山矶加州大学，医学院研究员

具体报名方法

有兴趣者请填写听课申请表,内容包括姓名,单位,职务职称,对该课程的学习与研究基础,联系方式等。免收学费，交通费、食宿费自理，如需听课资料，届时酌情收取工本费。莲花山研究院有非常好的食宿条件（具体见莲花山风景和住宿信息）。食宿费用来源一栏中需填自费或公费，如公费报销，需得到公费支出的负责人同意。填好听课申请表请发Email或传真传到龙星计划办公室张榕老师rzhang@ict.ac.cn，同时请抄送至研究院办公室夏文华wenhuaxia@gmail.com。

龙星计划课程官方网站：dragonstar.ict.ac.cn

食宿标准：

所有报名参加课程者一律在研究院进餐，标准为每人每天25元。住宿可选择研究院公寓（双人间，每人每天收费50元，仅限30人）或莲花山大门处的贵宾楼大厦（步行至研究院5分钟，双人间，每人每天60元或每标准间120元）。

其它信息诸如地理交通等请登录研究院网站查询：http://www.lotushill.org

 

授课人员和课程安排

科目1 Modeling, learning and conceptualizing visual patterns

Lecture Song-chun Zhu, Statistics and Computer Science, UCLA （Ph.D, Harvard)

       视觉统计建模与学习

主讲： 朱松纯，美国洛山矶加州大学统计系与计算机系教授 （美国哈佛大学博士)

10 topics：

Lecture 1 Modeling, learning and conceptualizing visual patterns

(The pursuit of descriptive and generative models, statistical observation in natural images)

Lecture 2 Constructing implicit and explicit manifolds

(The pursuit of descriptive and generative models)

Lecture 3 Minimax entropy learning for descriptive models

(Advanced Markov random fields and graphical models)

Lecture 4 Minimax entropy learning for generative models

(Clustering, epsilon-ball cover, and knowledge discovery)

Lecture 5 Mixing SCFG and Bi-gram

(A model integrating descriptive and generative components)

Lecture 6 Primal sketch: mixing structures and textures

(The representation for early vision)

Lecture 7 Scene representation with attribute graph grammar

Lecture 8  Learning  object categories with And-Or graph models 

Lecture 9 Information scaling, perceptual scale space, and transition of statistical regimes

Lecture 10  Recent advances in image parsing 

科目2 Psychophysical and cognitive Aspects of Vision

Lecturer: Hongjing Lu, Psychology and Statistics, UCLA (Ph.D in Psychology, UCLA)

视觉问题中的心理和认知科学研究

主讲： 吕宏静，美国洛山矶加州大学心理系与统计系博士后 (美国洛山矶加州大学博士)

5 topics：

Lecture 1: An introduction to vision science

      Most of us take completely for granted our ability to understand the world through seeing. However, we do not fully understand how humans are able quickly and effortlessly to perceive meaningful, coherent, three-dimensional scenes. I will demonstrate that perception is not a clear window onto reality, but rather an actively constructed, meaningful model of the world. Examples include phenomena related to adaptation, illusions, ambiguous figures, visual completion, and visual categorization. I will then describe four stages of visual perception—the image-based, surface-based, object-based, and category-based stages of perception—and two metaphorical information processing “directions”—bottom-up and top-down.

Lecture 2: Connecting the statistics in images with vision theories

      I will first present computational evidence on the basis of the work by Olshausen and Field (1998, 2004) to demonstrate the connection between the statistics in natural images and neural encoding of image-based information. Second, I will present the texton theory developed by Julesz (1981) and related experimental evidence.

Lecture 3: From natural statistics to generic priors

      This lecture will illustrate how prior knowledge affects human perception. This prior knowledge, which in its most general form is termed generic priors, is consistent with the statistics of natural scenes. Three studies will be presented, illustrating generic priors related to light-from-above (Kersten, Adelson, etc.), non-accidental viewpoints (Freeman), and slow-and-smooth motion (Weiss).

Lecture 4: Bayesian inference in cognitive science

      This lecture will illustrate the use of a Bayesian framework in the context of cognitive research investigating how humans draw strong inferences from noisy and sparse data.? I will present computational models that have been developed for two different domains in which inference is critical—motion perception and causal reasoning. The picture emerging from this work is that a key basis for human inference is the use of generic priors—tacit general assumptions people make about the way the world works, which then guide their learning and inference from observed data. I will sketch broader implications, including future directions for applying Bayesian modeling.

Lecture 5: Review of object recognition

      I will review the literature concerning object recognition from three perspectives—psychophysics, physiology, and computation. I will further discuss the strengths and weaknesses of two main theories, template matching and structural description.

科目3 Visual Motion Analysis

Lecture Ying Wu, EECS, Northwestern Universty (Ph.D, UIUC)

运动视觉分析

主讲： 吴郢, 美国西北大学电子工程与计算机科学系教授 (美国伊利诺伊大学博士)

5 topics (In about 10 lectures)：

Topic 1 Estimating 2D and 3D Motion from Image Sequences

(Motion field and optical flow Aperture problem Locas-Kanade’s method Dense flow Horn-Schunck’s method Parametric flow Robust flow computation Flow-based motion segmentation Generalized PCA Flow-based 3D motion analysis Direct methods Ego motion Parallax Multiple view motion analysis etc.)

Topic 2 Differential Motion Analysis

(Formulation Singularities Kernel-based methods Mean-shift tracking Support vector machine tracking Multiple kernel tracking Multiple collaborative kernel tracking, etc.)

Topic 3 Sequential Monte Carlo Motion Analysis

(Sequential Monte Carlo Particle filtering Factored sampling Importance sampling Markov chain Monte Carlo Metropolis-Hastings etc.)

Topic 4 Capturing Articulated Motion

 (Twists and exponential maps Markov random field Markov network Bayesian inference Belief propagation non-parametric BP Probabilistic variational analysis Mean-field variation Mean-field Monte Carlo Variational maximum a posteriori etc.)

Topic 5 Tracking Multiple Targets

(Coalescence Occlusions Data association Joint state space Probabilistic data association filtering Joint probabilistic data association filtering Decentralized tracking Tracking a variable number of targets MCMC tracking RJ-MCMC tracking etc.)

科目4 Stochastic computing and machine learning

      Lecture Zhuowen Tu, Lab of Neuro Im, aging, School of Medicine (Ph.D, Ohio State)

        随机计算和机器学习

主讲： 屠卓文,美国洛山矶加州大学医学院研究员 (美国俄亥俄大学博士)

5 topics：

Lecture 1 Computing discriminative models (Decision tree, Support Vector machine, C4.5, Bagging, AdaBoost, RealBoost)

Lecture 2 Modeling hierarchical priors (Dirichlet process, Chinese restaurant process, scene and object modeling)

Lecture 3 MCMC computing (Gibbs sampler, MCMC convergence analysis, Exact sampling)

Lecture 4 Data-driven Markov chain Monte Carlo (Image segmentation/parsing, Swendsen-Wang, Swnedsen-Wang Cuts and its generalization)

Lecture 5 Discriminative and generative computing for medical image analysis (Object detection, brain image segmentation, cortical structure detection)