| 3D Free-Form Object
Recognition in Range Images using Local Surface Patches
Presented by: Hui Chen
Recently, researches have been using surface signatures to
represent the shape of objects and recognize free-form objects
in range images. Surface signature is the local surface descriptor
that encodes the geometric properties of the neighborhood
of the point. Since it is local and could deal with the occlusion
and clutter well, it is becoming more and more attractive.
I'll talk about some typical papers using surface signatures.
But I think they have two disadvantages:1) They generate surface
signatures for every surface point during model building process.
2)Matching time increases linearly with the increase of the
model library. In our approach, we extract feature points
and define the local surface patche as feature point and its
neighbors. For every local surface patch, we compute the shape
indexes and normal angles between feature point and its neighbors.
In order to deal with large databases and speed up the search
process, we use mean and std of the shape index to hash a
table. Our approach has two steps. One is called off-line
processing which stores model information into a hashtable;
the other one is on-line recognition which select candidate
models by casting votes to the hashtable. Verification is
done by estimating the transformation and aligned the model
surface with scene surface.