From known to the unknown: Transferring knowledge to answer questions about novel visual and semantic concepts

摘要

Current Visual Question Answering (VQA) systems can answer intelligent questions about 'known' visual content. However, their performance drops significantly when questions about visually and linguistically 'unknown' concepts are presented during inference ('Open-world' scenario). A practical VQA systemshould be able to dealwith novel concepts in real world settings. To address this problem, we propose an exemplar-based approach that transfers learning (i.e., knowledge) from previously 'known' concepts to answer questions about the 'unknown'. We learn a highly discriminative joint embedding (JE) space, where visual and semantic features are fused to give a unified representation. Once novel concepts are presented to the model, it looks for the closest match from an exemplar set in the JE space. This auxiliary information is used alongside the given Image-Question pair to refine visual attention in a hierarchical fashion. Our novel attention model is based on a dual-attention mechanismthat combines the complementary effect of spatial and channel attention. Since handling the high dimensional exemplars on large datasets can be a significant challenge, we introduce an efficientmatching scheme that uses a compact feature description for search and retrieval. To evaluate ourmodel, we propose a newdataset for VQA, separating unknown visual and semantic concepts fromthe training set. Our approach shows significant improvements over state-of-the-art VQA models on the proposed Open-World VQA dataset and other standard VQA datasets. (c) 2020 Elsevier B.V. All rights reserved.