940 likes | 1.1k Views
The Odyssey of: Shape Matching with Ordered Boundary Point Shape Contexts Using a Least Cost Diagonal Method. By Dr. Carl E Abrams Jan 26th, 2008. Odyssey: a long wandering and eventful journey Or If we knew what we were doing, it wouldn't be called research, would it? – A Einstein.
E N D
The Odyssey of:Shape Matching with Ordered Boundary Point Shape Contexts Using a Least Cost Diagonal Method By Dr. Carl E Abrams Jan 26th, 2008
Odyssey: a long wandering and eventful journey Or If we knew what we were doing, it wouldn't be called research, would it? – A Einstein
Agenda • Topic Selection • The Eventful Journey • Elation • Mild Desperation • Elation • Beyond Desperation • Relief (Success) • Discussion / Lessons Learned
Topic Selection or… • First Idea: Requirements Engineering and Business Process Modeling • I despised the topic area
Topic Selection or…Roots • Identification of Automotive Vehicles Using Semantic Feature Extraction (Dec 2004) Elation!
Dissertation TimelineCarl E. Abrams Draft Chap 4&5 Now! Defense Draft Chap 3&4 9/03 1/04 1/05 1/06 Draft Idea Paper Complete Draft Proposal Advisor Selection Final Manuscript and Paper Complete Dissertation Proposal First Paper at Pace Day Committee Formation Final Draft Chap 1-3 Draft Chap 1&2
Semantic Geometric Features: A Preliminary Investigation of Automobile Identification Carl E. Abrams Sung-Hyuk Cha, Michael Gargano, and Charles Tappert
Agenda • Overview of the Problem • The Experiments • Results • Going Forward
Overview • Object recognition remains a hard problem • The human mind uses shapes to recognize objects • Can semantic features defined by their shapes be more effective in the recognition and identification of objects?
The Experiments • 10 test images of cars • Directly form the manufactures websites • Images were restricted to side views of the cars taken from 90 degrees • All 2005 models • Feature vectors calculated/measured from the images
Experiments used Euclidean Distance as the Measure the xi and ti are measurements from two different vehicles
c b a Experiments used Euclidean Distance as the Measure (x2,y2) c = (a2+b2)1/2 (x1,y1) c = ((x1-x2)2+(y1-y2)2)1/2
Challenge: Determine the qualitative ability of the feature vectors to separate the vehicles • Within each experiment compute the distance of each vehicle from all the others • Evenly divide the measures into 5 bins • Observe the distribution of the measures
Distance Matrix – Semantic Features Honda Civic Honda Accord Mazda 3 Mazda 6 Porsche Carerra Toyota Camry Toyota Celica Toyota Corolla Toyota Echo VW Passat Honda Civic Honda Accord Mazda 3 Mazda 6 Porsche Carerra Toyota Camry Toyota Celica Toyota Corolla Toyota Echo VW Passat
End of First Dead End Mild Desperation
Dissertation TimelineCarl E. Abrams Mild Desperation sets in Draft Chap 4&5 Now! Defense Draft Chap 3&4 9/03 1/04 1/05 1/06 Draft Idea Paper Complete Draft Proposal Advisor Selection Final Manuscript and Paper Complete Dissertation Proposal First Paper at Pace Day Committee Formation Final Draft Chap 1-3 Draft Chap 1&2
Shape Contexts • Shape Contexts are a novel shape descriptor introduced be Belongie[1] • Describes a shape by quantifying each point on the boundary of a shape by its relationship to all the on the boundary points on the shape • Compares shapes by comparing shape contexts [1] S. Belongie, "Image segmentation and shape matching for object recognition," vol. PhD, 2000, pp. 60.
Shape Contexts-Comparing Known CHI^2 Test where K= # of bins, g is unknown histogram and h is the known histogram Known Unknown Known
Progress ReportThe Role of Semantic Featuresin Automobile Identificationas of December 10th , 2005Carl E. Abrams
Agenda • Review of Topic and Approach (Elevator Pitch) • Summary Results to Date • Next Steps
Review of Topic and Approach • Develop an image segmentation and feature extraction / classification scheme for automobiles that employs the shapes of “semantic” parts and their geometric relationships. • Semantic features are the shapes of : windows, doors, front and rear quarter panels.
Review of Topic and Approach • Approach: • Develop a test database of vehicles by collecting side images • Develop/beg borrow or steal/software to interactively extract the shape and geometric information from the images • Work through all the classification test database • In parallel continue building the master database • Develop, test and compare segmentation / extraction method for semantic shapes
Preliminary Results-Test DBAs of Oct 15: All Ford models back to 1990 ~ 50 images As of Nov 12: Acura, Audi, Chrylser, Dodge, Ford, Honda, Mercury, Nissan, Pontiac, Saab, Saturn, Toyota, Volvo, VW models back to 1990 ~125 images
Preliminary Results-Image Segmentation Software • Modified CTMRedit: a matlab GUI for viewing, segmenting, and interpolating CT and MRI Images • Written by Mark Hasegawa-Johnson and Jul Cha • Simplified GUI and added capability to store shapes specific to vehicle identification
Preliminary Results • Create a feature vector that allows the comparison of one shape to another: Vector_Known(ws1,ws2,ws3,ws4,ds1,ds2,body shape) Vector_Unknown(ws1,ws2,ws3,ws4,ds1,ds2,body shape) Feature vector depends on shape descriptor in this case “Shape Contexts”
Preliminary Results – Shape Contexts • How to effectively describe a shape? r o
o Preliminary Results – Shape Contexts • How to effectively describe a shape? r r (5 bins) O (12 Bins)
Preliminary Results – Shape Contexts • How to effectively describe a shape? Develop the Shape Context histograms for every point on the shape
Preliminary Results – Distance Between Shape Contexts • How to effectively describe a shape? Known CHI^2 Test where K= # of bins, g is unknown histogram and h is the known histogram Each shape has 128 points, creates a 128x128 cost matrix Known Unknown Known
Preliminary Results – Distance Between Shape Contexts • What is the best fit (minimum cost) to align all the points? • The Assignment Problem – Hungarian Method for bi-partite matching problem We will be working with the following problem: assign n = 9 candidates to n=9 jobs to minimize the total salary cost paid by the department. The individual salaries of each candidate at each job position depend on their qualification and are given by the cost matrix (in $ per hour): If we start with the position of Administrator and assign it to Alex (he gets the minimal salary for this position), then we assign the position of Secretary to Chair to Lois (he gets the minimal salary for this position), and so on, up to the position of Typist, then the assignment is given by the assignment matrix:
Preliminary Results • Experimental Setup – Run 5 test cars against known database of 50 cars • Test cars re-segmented form known database • Plot out matches based on Euclidean Distance • Repeat by adding more cars of a different manufacturer to known DB
Preliminary Results Unknown: 2003FordMustang2DGT
Preliminary Results Unknown: 2003FordMustang2DGT-Volvos Added to Known DB
Preliminary Results Unknown: 2004FordFocus4DZX5.dh1
Preliminary Results Unknown: 2004FordFocus4DZX5.dh1-Volvos Added to Known DB
Preliminary Results Unknown: 2004FordTaurus4DSES.dh1
Preliminary Results Unknown: 2004FordTaurus4DSES.dh1-Volvos Added to Known DB
Preliminary Results Unknown: 2005FordMustangCoupe2DV6Deluxe.dh1