Graph-based selective rank fusion for unsupervised image retrieval

Valem, Lucas Pascotti [UNESP]; Pedronette, Daniel Carlos Guimarães [UNESP]

Graph-based selective rank fusion for unsupervised image retrieval

Data

2020-07-01

Autores

Valem, Lucas Pascotti

Pedronette, Daniel Carlos Guimarães

Tipo

Artigo

Resumo

Nowadays, there is a great variety of visual features available for image retrieval tasks. While fusion strategies have been established as a promising alternative, an inherent difficulty in unsupervised scenarios is the task of selecting the features to combine. In this paper, a Graph-based Selective Rank Fusion is proposed. The graph is used to represent the effectiveness estimation of features and the complementarity among them. The selected combinations are defined by the Connected Components of the graph. High-effective retrieval results were achieved through a comprehensive experimental evaluation considering different public datasets, dozens of features and comparisons with related methods. Relative gains up to +54.73% were obtained in relation to the best isolated feature.