Announcements. Recognition III. A Rough Recognition Spectrum. Projection, and reconstruction. Face detection using distance to face space

Similar documents
Eigenimages. Digital Image Processing: Bernd Girod, 2013 Stanford University -- Eigenimages 1

Eigenimages. Digital Image Processing: Bernd Girod, Stanford University -- Eigenimages 1

Fundamentals of Media Processing. Shin'ichi Satoh Kazuya Kodama Hiroshi Mo Duy-Dinh Le

Dimensionality Reduction PCA

Pattern Recognition Systems Lab 1 Least Mean Squares

Ones Assignment Method for Solving Traveling Salesman Problem

Lighting and Shading. Outline. Raytracing Example. Global Illumination. Local Illumination. Radiosity Example

Announcements. Recognition I. Optical Flow: Where do pixels move to? dy dt. I + y. I = x. di dt. dx dt. = t

Improving Template Based Spike Detection

EECS 442 Computer vision. Multiple view geometry Affine structure from Motion

3D Model Retrieval Method Based on Sample Prediction

Image Segmentation EEE 508

Bezier curves. Figure 2 shows cubic Bezier curves for various control points. In a Bezier curve, only

Administrative UNSUPERVISED LEARNING. Unsupervised learning. Supervised learning 11/25/13. Final project. No office hours today

SD vs. SD + One of the most important uses of sample statistics is to estimate the corresponding population parameters.

CSCI 5090/7090- Machine Learning. Spring Mehdi Allahyari Georgia Southern University

A New Morphological 3D Shape Decomposition: Grayscale Interframe Interpolation Method

EECS 442 Computer vision. Multiple view geometry Affine structure from Motion

COMP 558 lecture 6 Sept. 27, 2010

Performance Plus Software Parameter Definitions

The isoperimetric problem on the hypercube

Improving Face Recognition Rate by Combining Eigenface Approach and Case-based Reasoning

1 Graph Sparsfication

DATA MINING II - 1DL460

Octahedral Graph Scaling

Diego Nehab. n A Transformation For Extracting New Descriptors of Shape. n Locus of points equidistant from contour

Polynomial Functions and Models. Learning Objectives. Polynomials. P (x) = a n x n + a n 1 x n a 1 x + a 0, a n 0

Neuro Fuzzy Model for Human Face Expression Recognition

Our Learning Problem, Again

Normal Distributions

Math 10C Long Range Plans

Shape Completion and Modeling of 3D Foot Shape While Walking Using Homologous Model Fitting

Dynamic Programming and Curve Fitting Based Road Boundary Detection

Copyright 2016 Ramez Elmasri and Shamkant B. Navathe

Accuracy Improvement in Camera Calibration

( n+1 2 ) , position=(7+1)/2 =4,(median is observation #4) Median=10lb

Carnegie Mellon University

EE 584 MACHINE VISION

Alpha Individual Solutions MAΘ National Convention 2013

Image based Cats and Possums Identification for Intelligent Trapping Systems

Convex hull ( 凸殻 ) property

LDA-based Non-negative Matrix Factorization for Supervised Face Recognition

Learning to Shoot a Goal Lecture 8: Learning Models and Skills

localization error 1st pc pc 3 pc x2=

Real-Time Secure System for Detection and Recognition the Face of Criminals

Dimension Reduction and Manifold Learning. Xin Zhang

Creating Exact Bezier Representations of CST Shapes. David D. Marshall. California Polytechnic State University, San Luis Obispo, CA , USA

EigenFairing: 3D Model Fairing using Image Coherence

1.8 What Comes Next? What Comes Later?

9.1. Sequences and Series. Sequences. What you should learn. Why you should learn it. Definition of Sequence

arxiv: v2 [cs.ds] 24 Mar 2018

UNIT 4 Section 8 Estimating Population Parameters using Confidence Intervals

Consider the following population data for the state of California. Year Population

Elementary Educational Computer

Mobile terminal 3D image reconstruction program development based on Android Lin Qinhua

VALIDATING DIRECTIONAL EDGE-BASED IMAGE FEATURE REPRESENTATIONS IN FACE RECOGNITION BY SPATIAL CORRELATION-BASED CLUSTERING

Image Analysis. Segmentation by Fitting a Model

A SOFTWARE MODEL FOR THE MULTILAYER PERCEPTRON

EM375 STATISTICS AND MEASUREMENT UNCERTAINTY LEAST SQUARES LINEAR REGRESSION ANALYSIS

V.T. Chow, Open Channel Hydraulics, 1959 problem 9-8. for each reach computed in file below and placed here. = 5.436' yc = 2.688'

. Perform a geometric (ray-optics) construction (i.e., draw in the rays on the diagram) to show where the final image is formed.

The Closest Line to a Data Set in the Plane. David Gurney Southeastern Louisiana University Hammond, Louisiana

Numerical Methods Lecture 6 - Curve Fitting Techniques

South Slave Divisional Education Council. Math 10C

Wavelet Transform. CSE 490 G Introduction to Data Compression Winter Wavelet Transformed Barbara (Enhanced) Wavelet Transformed Barbara (Actual)

Eigendecomposition of Correlated Images Characterized by Three Parameters

Arithmetic Sequences

Lecture 18. Optimization in n dimensions

Counting Regions in the Plane and More 1

Lecture 2: Spectra of Graphs

15 UNSUPERVISED LEARNING

Describing data with graphics and numbers

The Virtual Point Light Source Model the Practical Realisation of Photometric Stereo for Dynamic Surface Inspection

IMP: Superposer Integrated Morphometrics Package Superposition Tool

Chapter 3: Introduction to Principal components analysis with MATLAB

CS 111: Program Design I Lecture 15: Objects, Pandas, Modules. Robert H. Sloan & Richard Warner University of Illinois at Chicago October 13, 2016

Protected points in ordered trees

1. Introduction o Microscopic property responsible for MRI Show and discuss graphics that go from macro to H nucleus with N-S pole

Lecture 1: Introduction and Strassen s Algorithm

Nonlinear Mean Shift for Clustering over Analytic Manifolds

Copyright 2016 Ramez Elmasri and Shamkant B. Navathe

New HSL Distance Based Colour Clustering Algorithm

Theory of Fuzzy Soft Matrix and its Multi Criteria in Decision Making Based on Three Basic t-norm Operators

Texture Analysis and Indexing Using Gabor-like Hermite Filters

BASED ON ITERATIVE ERROR-CORRECTION

Behavioral Modeling in Verilog

Hand Gesture Recognition for Human-Machine Interaction

Structuring Redundancy for Fault Tolerance. CSE 598D: Fault Tolerant Software

BOOLEAN MATHEMATICS: GENERAL THEORY

RADIAL BASIS FUNCTION USE FOR THE RESTORATION OF DAMAGED IMAGES

Soft Computing Based Range Facial Recognition Using Eigenface

ON THE QUALITY OF AUTOMATIC RELATIVE ORIENTATION PROCEDURES

Cubic Polynomial Curves with a Shape Parameter

LU Decomposition Method

Evaluation of Support Vector Machine Kernels for Detecting Network Anomalies

Solutions to Final COMS W4115 Programming Languages and Translators Monday, May 4, :10-5:25pm, 309 Havemeyer

Evaluation scheme for Tracking in AMI

Designing a learning system

Using the Keyboard. Using the Wireless Keyboard. > Using the Keyboard

Designing a learning system

Transcription:

Aoucemets Assigmet 5: Due Friday, 4:00 III Itroductio to Computer Visio CSE 52 Lecture 20 Fial Exam: ed, 6/9/04, :30-2:30, LH 2207 (here I ll discuss briefly today, ad will be at discussio sectio tomorrow for first 45 miutes. A Rough Spectrum Virtual Ciematography: Makig 'he Matrix' Sequels George Borshukov VFX echology Supervisor, ESC Etertaimet Friday, Jue 4, 2004 :00 p.m. to 2:30 p.m. [Pizza luch will precede the evet from oo to p.m.] Mai Auditorium, Sa Diego Supercomputer Ceter he presetatio will cover the key techologies that had to be developed ad deployed to create the sythetic huma sequeces i the Matrix sequels icludig Uiversal Capture - image-based facial aimatio, realistic huma face rederig, ad use of measured BRDF i film productio. It will also feature a breakdow of he Superpuch shot (pictured above from "he Matrix Revolutios" (the bullet time puch that Neo delivers to Aget Smith durig the film's last face-off. his difficult, importat, expesive, ad challegig shot was etirely computer geerated ad showcased the techological developmets of 3.5+ years at their best by showig a full-frame close-up of a kow huma actor. Appearace-Based (Eigeface, Fisherface Shape Cotexts Local Features + Spatial Relatios Geometric Ivariats Aspect Graphs Icreasig Geerality 3-D Model-Based Image Abstractios/ Volumetric Primitives Fuctio Projectio, ad recostructio A -pixel image x R ca be projected to a low-dimesioal feature space y R m by y = x From y R m, the recostructio of the poit is y he error of the recostructio is: x- x Face detectio usig distace to face space Sca a widow ω across the image, ad classify the widow as face/ot face as follows: Project widow to subspace, ad recostruct as described earlier. Compute distace betwee ω ad recostructio. Local miima of distace over all image locatios less tha some treshold are take as locatios of faces. Repeat at differet scales. Possibly ormalize widows itesity so that ω =.

Sigular Value Decompositio Ay m by matrix A may be factored such that A = UΣV [m x ] = [m x m][m x ][ x ] U: m by m, orthogoal matrix Colums of U are the eigevectors of AA V: by, orthogoal matrix, colums are the eigevectors of A A Σ: m by, diagoal with o-egative etries (σ, σ 2,, σ s with s=mi(m, are called the called the sigular values Sigular values are the square roots of eigevalues of both AA ad A A & Colums of U are correspodig Eigevectors!! Performig PCA with SVD Sigular values of A are the square roots of eigevalues of both AA ad A A & Colums of U are correspodig Eigevectors Ad a iai = [ a a2 L a ][ a a2 L a ] = AA i= Covariace matrix is: Σ = i= r r r r ( x i µ ( x µ So, igorig / subtract mea image µ from each iput image, create data matrix, ad perform (thi SVD o the data matrix. i Result of SVD algorithm: σ σ 2 σ s PCA & Fisher s Liear Discrimiat PCA PCA (Eigefaces χ χ 2 PCA = arg max S Maximizes projected total scatter Fisher s Liear Discrimiat FLD fld SB = arg max S Maximizes ratio of projected betwee-class to projected withi-class scatter Variability: Camera positio Illumiatio Iteral parameters ithi-class variatios A example: surfaces of first 3 coefficiets Parameterized Eigespace

Basic ideas i classifiers Bayesia Classificatio Discussed o blackboard, but slides may be helpful Loss some errors may be more expesive tha others e.g. a fatal disease that is easily cured by a cheap medicie with o side-effects -> false positives i diagosis are better tha false egatives e discuss two class classificatio: L(->2 is the loss caused by callig a 2 otal risk of usig classifier s Basic ideas i classifiers Geerally, we should classify as if the expected loss of classifyig as is better tha for 2 gives Some loss may be ievitable: the miimum risk (shaded area is called the Bayes risk if 2 if Crucial otio: Decisio boudary poits where the loss is the same for either case Example: kow distributios Fidig a decisio boudary is ot the same as modellig a coditioal desity. pxk ( = 2π p 2 Σ 2 exp 2 x µ k ( Σ ( x µ k Assume ormal class desities, p-dimesioal measuremets with commo (kow covariace ad differet (kow πmeas k Class priors are Ca igore a commo factor i posteriors - importat; posteriors are the: pk ( x ( π k p 2 Σ 2 exp 2π 2 x µ k ( Σ ( x µ k

Classifier boils dow to: choose class that miimizes: Mahalaobis distace δ( x, µ k 2 2 log π k where δ x, µ ( k = x µ k ( Σ x µ ( k because covariace is commo, this simplifies to sig of a liear expressio (i.e. Vorooi diagram i 2D for Σ=I 2 Fidig ski Ski has a very small rage of (itesity idepedet colours, ad little texture Compute a itesity-idepedet colour measure, check if colour is i this rage, check if there is little texture (media filter See this as a classifier - we ca set up the tests by had, or lear them. get class coditioal desities (histograms, priors from data (coutig Classifier is Receiver Operatig Curve Figure from Statistical color models with applicatio to ski detectio, M.J. Joes ad J. Rehg, Proc. Computer Visio ad Patter, 999 copyright 999, IEEE Figure from Statistical color models with applicatio to ski detectio, M.J. Joes ad J. Rehg, Proc. Computer Visio ad Patter, 999 copyright 999, IEEE Appearace-Based Visio: Lessos Stregths Posig the recogitio metric i the image space rather tha a derived represetatio is more powerful tha expected. Modelig objects from may images is ot ureasoable give hardware developmets. he data (images may provide a better represetatios tha abstractios for may tasks. Appearace-Based Visio: Lessos eakesses Segmetatio or object detectio is still a issue. o trai the method, objects have to be observed uder a wide rage of coditios (e.g. pose, lightig, shape deformatio. Limited power to extrapolate or geeralize (abstract to ovel coditios.

Model-Based Visio A Rough Spectrum Appearace-Based (Eigeface, Fisherface Shape Cotexts Geometric Ivariats Image Abstractios/ Volumetric Primitives Give 3-D models of each object Detect image features (ofte edges, lie segmets, coic sectios Establish correspodece betwee model &image features Estimate pose Cosistecy of projected model with image. Local Features + Spatial Relatios Aspect Graphs 3-D Model-Based Fuctio by Hypothesize ad est Geeral idea Hypothesize object idetity ad pose Recover camera parameters (widely kow as backprojectio Reder object usig camera parameters Compare to image Issues where do the hypotheses come from? How do we compare to image (verificatio? Simplest approach Costruct a correspodece for all object features to every correctly sized subset of image poits hese are the hypotheses Expesive search, which is also redudat. Correspodeces betwee image features ad model features are ot idepedet. A small umber of correspodeces yields a camera matrix --- the others correspodeces must be cosistet with this. Pose cosistecy Strategy: Geerate hypotheses usig small umbers of correspodeces (e.g. triples of poits for a calibrated perspective camera, etc., etc. Backproject ad verify Scee Iterpretatio he Swig Fragoard, 766 Fial Exam Closed book Oe cheat sheet Sigle piece of paper, hadwritte, o photocopyig, o physical cut & paste. you ca start with sheet from the midterm, if you wat. hat to study Basically material preseted i class, ad supportig material from text If it was i text, but NEVER metioed i class, it is very ulikely to be o the exam Questio style: Short aswer Some loger problems to be worked out.