Learning context-sensitive similarity by shortest path propagation

In this paper, we introduce a novel shape/object retrieval algorithm shortest path propagation (SSP). Given a query object q and a target database object p, we explicitly find the shortest path between them in the distance manifold of the database objects. Then a new distance measure between q and p is learned based on the database objects on the shortest path to replace the original distance measure. The promising results on both MEPG-7 shape dataset and a protein dataset demonstrate that our method can significantly improve the ranking of the object retrieval.

Research highlights
► GP assumes the similarity between query and target is affected by all other objects.
► We argue it is only affected by a few contextual reference objects.
► These reference objects are the nodes in the shortest path between them.
► Our SSP propagates similarity from query to target along the shortest path.