CSE 803 Fall 20091 2D matching part 2 Review of alignment methods and errors in using them Introduction to more 2D matching methods.

CSE 803 Fall 2009 1

2D matching part 2

• Review of alignment methods and

errors in using them

• Introduction to more 2D matching

methods

CSE 803 Fall 2009 2

Review of roadmap: algorithms to control matching

CSE 803 Fall 2009 3

Rigid transformation review

CSE 803 Fall 2009 4

Affine includes scaling and shear

CSE 803 Fall 2009 5

Problems with error Least squares fitting uses n >> 3 point

pairs Significantly reduces error across field Will still be thrown off by outliers * can throw out pairs with high error and then refit * can set the “weight” of any pair to be inversely proportional to error squared

CSE 803 Fall 2009 6

Sources of error

* Wrong matching in the pair of points yields outlier

CSE 803 Fall 2009 7

2-Point alignment error due to error in locations of Q1, Q2

Plastic slides can actually be overlaid for better viewing.

CSE 803 Fall 2009 8

Remove outlier and refit

Plastic slides show concept better.

CSE 803 Fall 2009 9

Sometimes a halucination

6 points match, but the objects do not. Can verify using more model points.

CSE 803 Fall 2009 10

Local Feature Focus Method (Bolles)

CSE 803 Fall 2009 11

Local focus feature matching Local features tolerate occlusion

by other objects (binpicking problem)

Subgraph matching provides several features (distances, angles, connections, etc.)

Method can be used to support different higher level strategies and alignment parameters

CSE 803 Fall 2009 12

Focus features matching attempts

CSE 803 Fall 2009 13

Pose clustering (generalized Hough transform)

Use m minimal sets of matching features, each just enough to compute alignment

Vector of alignment parameters is put as evidence into “parameter space”

When all m units of evidence computed, examine parameter space for cluster[s]

CSE 803 Fall 2009 14

Pose clustering

CSE 803 Fall 2009 15

Line segment junctions for matching

CSE 803 Fall 2009 16

“Abstract vectors” subtending detected junctions

Abstract vector with tail at T and tip at Y, or tail at L and tip at X

MAP IMAGE

CSE 803 Fall 2009 17

Parameter space resulting from 10 vector matches

Rotation, scale, translation computed as in single match alignment. Use the cluster center to estimate best alignment parameters.

CSE 803 Fall 2009 18

Detecting airplanes on airfield

CSE 803 Fall 2009 19

Airplane model of abstract vectors; detected image features

CSE 803 Fall 2009 20

Relational matching method

CSE 803 Fall 2009 21

Some relations between parts

CSE 803 Fall 2009 22

Recognition via consistent labeling

CSE 803 Fall 2009 23

Parts, labels, relations

CSE 803 Fall 2009 24

In a consistent labeling “image” parts relate as do “model” parts

CSE 803 Fall 2009 25

Distance relation often used

CSE 803 Fall 2009 26

What model labels apply to detected holes H1, H2, H3?

CSE 803 Fall 2009 27

Partial Interpretation Tree to find a distance consistent labeling

The IT shows matching attempts that can be tried using a backtracking algorithm. If a relation fails the algorithm tries a different branch.

CSE 803 Fall 2009 28

Detailed IT algorithm

Current matching pairs can be stored in the recursive stack. If a new pair is consistent with the previous pairs, continue forward; if not, then back up (and retract the recent pairing).

CSE 803 Fall 2009 29

CSE 803 Fall 2009 30

Discrete relaxation labeling constrains possible labels

A sometimes useful method that once drew much interest (see pubs by Rosenfeld, Zucker, Hummel, etc.) The Marr-Poggio stereo matching algorithm has the character of relaxation.

CSE 803 Fall 2009 31

Discrete relaxation labeling constrains possible labels

CSE 803 Fall 2009 32

Kleep matching via relaxation

CSE 803 Fall 2009 33

Removing a possible label for one part affects labels for related parts

CSE 803 Fall 2009 34

Relaxation labeling Can work truly in parallel Pairwise constraints are weaker than

what the IT method can check, so sometimes the IT must follow the relaxation method

There is “probabalistic relaxation” which changes probability of labels rather than just keeping or deleting them

Relaxation was once thought to model human visual processes.