HADA: A Graph-Based Amalgamation Framework in Image-text Retrieval

Abstract

Many models have been proposed for vision and language tasks, especially the image-text retrieval task. State-of-the-art (SOTA) models in this challenge contain hundreds of millions of parameters. They also were pretrained on large external datasets that have been proven to significantly improve overall performance. However, it is not easy to propose a new model with a novel architecture and intensively train it on a massive dataset with many GPUs to surpass many SOTA models already available to use on the Internet. In this paper, we propose a compact graph-based framework named HADA, which can combine pretrained models to produce a better result rather than starting from scratch. Firstly, we created a graph structure in which the nodes were the features extracted from the pretrained models and the edges connecting them. The graph structure was employed to capture and fuse the information from every pretrained model. Then a graph neural network was applied to update the connection between the nodes to get the representative embedding vector for an image and text. Finally, we employed cosine similarity to match images with their relevant texts and vice versa to ensure a low inference time. Our experiments show that, although HADA contained a tiny number of trainable parameters, it could increase baseline performance by more than 3.6% in terms of evaluation metrics on the Flickr30k dataset. Additionally, the proposed model did not train on any external dataset and only required a single GPU to train due to the small number of parameters required. The source code is available at https://github.com/m2man/HADA.