Unsupervised Similarity Learning through Rank Correlation and kNN Sets

摘要: The increasing amount of multimedia data collections available today evinces the pressing need for methods capable indexing and retrieving this content. Despite continuous advances in features representation models, to establish an effective measure comparing different objects still remains a challenging task. While supervised semi-supervised techniques made relevant on similarity learning tasks, scenarios where labeled are non-existent require strategies. In such situations, unsupervised has been established as promising solution, considering contextual information dataset structure computing new similarity/dissimilarity measures. This article extends recent algorithm that uses iterative re-ranking strategy take advantage k-Nearest Neighbors (kNN) sets rank correlation Two novel approaches proposed kNN their corresponding top-k lists. were validated conjunction with various measures, yielding superior effectiveness results comparison previous works. addition, we also evaluate ability method objects, conducting extensive experimental evaluation image video datasets.