Indexing Adaptive Similarity

When determining content-based similarity between two multimedia objects, the distance is evaluated on feature representations which aggregate the inherent properties of the multimedia objects. The conventional feature representations aggregate and store these properties in feature histograms, which can be compared by vectorial distances. Recent feature representations adaptively aggregate and store individual object properties in more flexible feature signatures, which can be compared by adaptive similarity measures, such as the quadratic form distance or the Earth mover's distance.However, as the signature-based distances are computationally expensive, their efficient employment into large-scale search retrieval systems remains a challenging issue. To achieve competitive results using the adaptive distances, SRG aims to investigate non-conventional indexing approaches, such as the ptolemaic indexing or synergistic adjustation of the similarity model.