Results

Shock-Graph Matching Results

• Shock-graph matching  gives intuitive correspondences for a variety of shapes

Robustness to Visual Transformations

• The medial axis is sensitive to boundary perturbations, which introduce spurious edges in the graph

• Typically tackled by regularization during the detection process or post-detection

• Here regularization is an integral part of the recognition process

• The shock graph inherently segments a shape into parts  and captures the hierarchical relationship between them

• Hence it is robust to changes which may occur in some of the parts

• When viewpoint is varied gradually, the spatial location and the shape of parts  typically changes gradually, and are handled by deform edit

• At certain views,  there is a sudden appearance or disappearance of a part, which are handled by the remaining edits

• Appearance of a part is handled by the splice edit

• A change in aspect is handled by the contract edit

• Leads to an aspect graph approach to 3D object recognition [Cyr:Kimia:ICCV01]

• Shadows and highlights often lead to segmentation errors in figure-ground segregation

•  Highlights tend to cause small changes in the boundary, but typically affect the shock-graph topology.

• Shadows are often more global in nature, but tend not to affect the shock-graph topology

• Shock-graph matching gives the intuitive correspondence by splicing out the occluded part

• Two types of occlusion are considered

• Occluder blends with the shape

• Occluder blends with the background

• Plot of edit cost in the presence of increasing  occlusion

• May fail if a significantly large portion of the shape is occluded

Indexing Results

• Two databases

• The first database (99 shapes) has nine categories (fish, rabbit, airplane, ``greeble'', tool, hand, doll, four-legged animal, and sea-animal) eleven shapes in each category

• Acknowledgements: Mike Tarr, Farzin Mokhtarian, Stan Sclaroff
• Shapes have variations  in form, as well as for occlusion, articulation, missing parts

• Each shape is matched against all others

• Results quantified in terms of top 10 matches from same category
 

	1	2	3	4	5	6	7	8	9	10
%	100	100	100	99	99	99	97	96	95	87

• Top three matches are always correct




• The second database (216 shapes) has eighteen categories with twelve shapes in each category

• Acknowledgements: MPEG-7 test database

• Top 11 matches are from same category

 

	1	2	3	4	5	6	7	8	9	10	11
%	100	100	100	99	97	99	96	96	95	91	80

• Top three matches are always correct