Feature triplets for object recognition
L. Zitnick
Matching primitives for feature-based object recognition should
be both
discrimitive between different objects, and spatially redicative of
neighboring features. While "bag of words" approaches rely only on
a
feature's discrimitive power, recent results have demonstrated that
modeling the spatial relationships of features can improve performance.
To achieve both of these goals, we propose using groups of three
features called triplets, instead of single features as our matching primitive.
We apply the triplet approach to two problems: object instance
recognition in large databases, and object category recognition. For
object instance recognition, we use an image-centric object model with
global spatial constraints. Given the added discrimitive power of
triplets, a sparse set of potential matches are found. The matches are
verified by enforcing their spatial consistency using a geometric
hashing technique.
For object category recognition, a model-centric approach is taken with local spatial constraints. Triplets from multiple training images of an object are combined into a single model based on appearance.