Recurrent Relational Memory Network for Unsupervised Image Captioning

TL;DR

This paper introduces R^2M, a memory-based network for unsupervised image captioning that avoids GANs, uses a two-stage memory mechanism for relational reasoning, and outperforms existing methods on benchmarks.

Contribution

The paper presents a novel R^2M network that replaces GANs with a memory-based approach for unsupervised image captioning, improving efficiency and performance.

Findings

R^2M outperforms state-of-the-art methods on benchmark datasets.

The proposed model has fewer parameters and higher computational efficiency.

It effectively encodes visual context and learns from textual corpora without supervision.

Abstract

Unsupervised image captioning with no annotations is an emerging challenge in computer vision, where the existing arts usually adopt GAN (Generative Adversarial Networks) models. In this paper, we propose a novel memory-based network rather than GAN, named Recurrent Relational Memory Network ($R^2M$). Unlike complicated and sensitive adversarial learning that non-ideally performs for long sentence generation, $R^2M$ implements a concepts-to-sentence memory translator through two-stage memory mechanisms: fusion and recurrent memories, correlating the relational reasoning between common visual concepts and the generated words for long periods. $R^2M$ encodes visual context through unsupervised training on images, while enabling the memory to learn from irrelevant textual corpus via supervised fashion. Our solution enjoys less learnable parameters and higher computational efficiency than GAN-based methods, which heavily bear parameter sensitivity. We experimentally validate the superiority of $R^2M$ than state-of-the-arts on all benchmark datasets.