Contour-Based Recognition We now have a set of contours and their - - PowerPoint PPT Presentation

47

Contour-Based Recognition

• We now have a set of contours and their

relations (proximal and parallel, since we can merge collinear contours into a single contour).

• These can be matched to a similar model

description, made up of model contours and model relations (or constraints).

• How do we match the two?

48

Interpretation Tree (IT) Search

• Introduced by Grimson and Lozano-Perez

(1984).

• Equivalent to a constraint satisfaction problem

(CSP), where model labels are assigned to image features subject to their satisfying model constraints.

• Equivalent to a subgraph isomorphism problem,

where an edge- and node-attributed model (sub)graph is matched to an image graph.

49

Example

M1 M2 M3 M4 M5 M6 Proximity (cotermination) Parallelism Object boundary Model: Scene: I1 I2 I3 I4 I5

50

Example

M1 M2 M3 M4 M5 M6

Model: Scene:

I1 I2 I3 I4 I5

I1 I2 I3 I4 I5

M1 M3 M6 M6 M4 M1 M4 M4 M3 M2 M5 M4 M3 M3 M2 M3 M4 M1 M4 M3 M2 M2 M2 M1 M3 M3 M1 M2 M5 M6 M5 M2 M1 M2 M5 M3 M5 M5 M4

… … … …

M2 M3

51

Spurious Features

• Spurious features need wildcard assignment.
• Increases complexity of search significantly.

M1 M2 M3 M4 M5 M6

Model: Scene:

I1 I2 I3 I4 I5

I5 I2

M1 M3 M6 M2 M5 M4 WC M6 M4 M3 M2 M5 M1

52

IT Epilogue

• Like any search problem, efficiency can be

gained through correct feature ordering, e.g., match most discriminative (least ambiguous) features first.

• Nature of features and constraints dictates

how categorical/exemplar the model is.

• The stronger the features/constraints, the

more aggressive the pruning, but the more brittle (exemplar-based) the model.