Modality-specific Cross-modal Similarity Measurement with Recurrent Attention Network

摘要

Nowadays, cross-modal retrieval plays an indispensable role to flexibly findinformation across different modalities of data. Effectively measuring thesimilarity between different modalities of data is the key of cross-modalretrieval. Different modalities such as image and text have imbalanced andcomplementary relationships, which contain unequal amount of information whendescribing the same semantics. For example, images often contain more detailsthat cannot be demonstrated by textual descriptions and vice versa. Existingworks based on Deep Neural Network (DNN) mostly construct one common space fordifferent modalities to find the latent alignments between them, which losetheir exclusive modality-specific characteristics. Different from the existingworks, we propose modality-specific cross-modal similarity measurement (MCSM)approach by constructing independent semantic space for each modality, whichadopts end-to-end framework to directly generate modality-specific cross-modalsimilarity without explicit common representation. For each semantic space,modality-specific characteristics within one modality are fully exploited byrecurrent attention network, while the data of another modality is projectedinto this space with attention based joint embedding to utilize the learnedattention weights for guiding the fine-grained cross-modal correlationlearning, which can capture the imbalanced and complementary relationshipsbetween different modalities. Finally, the complementarity between the semanticspaces for different modalities is explored by adaptive fusion of themodality-specific cross-modal similarities to perform cross-modal retrieval.Experiments on the widely-used Wikipedia and Pascal Sentence datasets as wellas our constructed large-scale XMediaNet dataset verify the effectiveness ofour proposed approach, outperforming 9 state-of-the-art methods.