Indoor Object Recognition f| 
• Objects are moveable 
• only qualitative context can be used 
e.g. telephone, computer monitor usually on a table 
• Matching needs to be local 
• Features such as lines allow too many matches 
such systems usually solve “pose estimation” problems 
• Grouped features are much more distinctive 
e.g. lines belonging to a single object 
• Higher level surface or volume descriptions are efficient 
Conclusion 
• Grouping and Matching are essential operations 
• Grouping and matching can be mutually supportive 
• Invariant, geometric properties are important in both 
• Complexity can be controlled by hierarchical processes 
• Inference or measurement of 3-D simplifies the processes 
can handle large object databases 
viewgraphs.fm—15 
8/4/95 
viewgraphs.fm—16 
8/4/95 
Nevatia - 18