Learning Two-Branch Neural Networks for Image-Text Matching Tasks

TL;DR

This paper introduces two neural network architectures for image-text matching, demonstrating improved accuracy in phrase localization and image-sentence retrieval tasks through novel training strategies and network designs.

Contribution

It proposes two distinct two-branch neural network structures with innovative training methods for enhanced image-text matching performance.

Findings

High accuracy in phrase localization on Flickr30K Entities

Effective bi-directional image-sentence retrieval on Flickr30K and MSCOCO

Improved neighborhood sampling enhances training effectiveness

Abstract

Image-language matching tasks have recently attracted a lot of attention in the computer vision field. These tasks include image-sentence matching, i.e., given an image query, retrieving relevant sentences and vice versa, and region-phrase matching or visual grounding, i.e., matching a phrase to relevant regions. This paper investigates two-branch neural networks for learning the similarity between these two data modalities. We propose two network structures that produce different output representations. The first one, referred to as an embedding network, learns an explicit shared latent embedding space with a maximum-margin ranking loss and novel neighborhood constraints. Compared to standard triplet sampling, we perform improved neighborhood sampling that takes neighborhood information into consideration while constructing mini-batches. The second network structure, referred to as a similarity network, fuses the two branches via element-wise product and is trained with regression loss to directly predict a similarity score. Extensive experiments show that our networks achieve high accuracies for phrase localization on the Flickr30K Entities dataset and for bi-directional image-sentence retrieval on Flickr30K and MSCOCO datasets.