Aligning the True Semantics: Constrained Decoupling and Distribution Sampling for Cross-Modal Alignment

TL;DR

This paper introduces a novel cross-modal alignment method called CDDS that decouples semantic and modality information in embeddings, using constrained dual-path UNet and distribution sampling to improve semantic consistency between vision and language.

Contribution

The paper proposes a new algorithm, CDDS, that effectively separates semantic from modality information and bridges modality gaps, advancing cross-modal alignment techniques.

Findings

Outperforms state-of-the-art methods by 6.6% to 14.2% on various benchmarks.

Uses a dual-path UNet for adaptive embedding decoupling.

Employs distribution sampling to address modality gaps.

Abstract

Cross-modal alignment is a crucial task in multimodal learning aimed at achieving semantic consistency between vision and language. This requires that image-text pairs exhibit similar semantics. Traditional algorithms pursue embedding consistency to achieve semantic consistency, ignoring the non-semantic information present in the embedding. An intuitive approach is to decouple the embeddings into semantic and modality components, aligning only the semantic component. However, this introduces two main challenges: (1) There is no established standard for distinguishing semantic and modal information. (2) The modality gap can cause semantic alignment deviation or information loss. To align the true semantics, we propose a novel cross-modal alignment algorithm via \textbf{C}onstrained \textbf{D}ecoupling and \textbf{D}istribution \textbf{S}ampling (CDDS). Specifically, (1) A dual-path UNet is introduced to adaptively decouple the embeddings, applying multiple constraints to ensure effective separation. (2) A distribution sampling method is proposed to bridge the modality gap, ensuring the rationality of the alignment process. Extensive experiments on various benchmarks and model backbones demonstrate the superiority of CDDS, outperforming state-of-the-art methods by 6.6\% to 14.2\%.