Deep Vision-Language Model for Efficient Multi-modal Similarity Search in Fashion Retrieval

Fashion multi-modal retrieval has been recently addressed using vision-and-language transformers. However, these models cannot scale in training time and memory requirements due to the quadratic attention mechanism. Moreover, they design the retrieval as a classification task, assigning a similarity score to pairs of text and images in input. Each query is thus resolved inefficiently by pairing it, at runtime, with every text or image in the entire dataset, precluding the scalability to large-scale datasets. We propose a novel approach for efficient multi-modal retrieval in the fashion domain that combines self-supervised pretraining with linear attention and deep metric learning to create a latent space where spatial proximity among instances translates into a semantic similarity score. Unlike existing contributions, our approach separately embeds text and images, decoupling them and allowing to collocate and search in the space, after training, even for new images with missing text and vice versa. Experiments show that with a single 12 GB GPU, our solution outperforms, both in efficacy and efficiency, existing state-of-the-art contributions on the FashionGen dataset. Our architecture also enables the adoption of multidimensional indices, with which retrieval scales in logarithmic time up to millions, and potentially billions, of text and images.