Visual Class Models


A visual class model is a description of the appearance of groups of objects such as horse, bicycle, butterfly such that individual instances can be recognised as belonging to the class.

Standard Computer Vision methods assume (tacitly or otherwise) that photographic appearance, because they rely on low variation in the description and train over photographs. When trained on other depictive styles, classification performance falls quite significantly.

We are interested in building visual class models that are robust to depictive style.

Q. Wu and P.Hall. Modelling Visual Objects Invariant to Depictive Style. BMVC 2013.

We input images showing objects is different depictive styles. We build a graph based model of the whole class. Nodes are labelled with prime shapes.

This process results in models, one per visual class of interest.

Testing shows we improve cross-domain from about 50% to 60%, compared to bog of words.