Cross-modal Retrieval for Knowledge-based Visual Question Answering

TL;DR

The paper tackles knowledge-based visual question answering for named entities (KVQAE) by addressing the retrieval bottleneck in unstructured knowledge bases. It leverages a multimodal dual encoder (CLIP) to model both mono-modal (IqIp) and cross-modal (IqTp) retrieval, and studies three training strategies—mono-modal, cross-modal, and hybrid—finding that combining the two yields the strongest performance. Across ViQuAE, InfoSeek, and Encyclopedic-VQA, cross-modal retrieval consistently outperforms mono-modal retrieval, and a hybrid approach achieves competitive results with billion-parameter models while remaining computationally efficient. By integrating DPR-based passage retrieval and Multi-passage BERT for answer extraction, the method demonstrates robust passage-level MRR improvements, though it also highlights KB incompleteness and a sizeable gap to oracle retrieval, indicating IR as the primary bottleneck. Overall, the work shows that cross-modal cues effectively bridge heterogeneous visual representations of entities and paves the way for retrieval-based KVQAE in more scalable and controllable KB settings.

Abstract

Knowledge-based Visual Question Answering about Named Entities is a challenging task that requires retrieving information from a multimodal Knowledge Base. Named entities have diverse visual representations and are therefore difficult to recognize. We argue that cross-modal retrieval may help bridge the semantic gap between an entity and its depictions, and is foremost complementary with mono-modal retrieval. We provide empirical evidence through experiments with a multimodal dual encoder, namely CLIP, on the recent ViQuAE, InfoSeek, and Encyclopedic-VQA datasets. Additionally, we study three different strategies to fine-tune such a model: mono-modal, cross-modal, or joint training. Our method, which combines mono-and cross-modal retrieval, is competitive with billion-parameter models on the three datasets, while being conceptually simpler and computationally cheaper.

Figures (3)

• Figure 1: Two visual questions from the ViQuAE dataset along with relevant visual passages from its Knowledge Base. The different types of mono- and cross-modal interactions studied are also shown for the second question. The acronyms of the interactions are composed of the letters T (Text), I (Image), q (question) and p (passage).
• Figure 2: Strengths and weaknesses of mono- and cross-modal retrieval exemplified through CLIP results (not fine-tuned) on ViQuAE's validation set.
• Figure 3: Passage-level MRR on the validation set of ViQuAE depending on the $\alpha_{\{T,I,C\}}$ hyperparameters.