Capturing Spatial Context in Images with a Relational Dictionary
Probabilistic generative models such as Probabilistic Latent Semantic Indexing (PLSI), Latent Dirichlet Allocation (LDA), and their variants have been used with some success on the task of object recognition in images. Recent approaches integrate spatial information into the generative process leading to complex hierarchical models. This work takes a simpler approach by encoding the neighborhood of a particular visual word in a relation dictionary. Thus, an image is represented as histogram over this relational dictionary, and standard topic modeling approaches can be applied. We examine three methods of inducing the neighborhood graph and compare them on an image recognition task using the LabelMe benchmark database. Results for the relational dictionary are on-par with a standard non-relational dictionary, but consistently poor, thereby suggesting a flaw in the experimental setup.
Downloads
 Final Paper |