QPert: Query Perturbation to improve shape retrieval algorithms

Although there is a wide range of shape descriptors available in the literature, most of them are restricted to a specific class of shapes and no one can achieve satisfactory shape retrieval results when used with different classes of shapes. Introducing new descriptors, improving, or merging existing descriptors are potential strategies for enhancing shape retrieval algorithms. In this paper, we propose a Query Perturbation-based (QPert) method for shape retrieval. QPert perturbs the query shape to create copies or clones that are closer than the query itself to the database shapes. Clones are created by adding a small noise to the coordinates of a randomly selected subset of mesh vertices or applying genetic operators between existing clones. A Genetic Algorithm (GA) gradually develops a population of clones so that the fittest clones get closer and closer to their most similar shapes in the database. The GA is implemented as a multiagent system (MAS) that enables any number of shape descriptors, classical or modern, to cooperate without the need for synchronization or direct communication between agents. Experimental results and comparisons demonstrate the advantages of this approach, regardless of the shape descriptors used.