Title: MMDG-Bench: A Benchmark for Multimodal Domain Generalization

Abstract

Multi-modal Domain Generalization (MMDG) aims to improve model robustness across unseen domains by capitalizing on the complementary nature of different modalities. While Multi-modal Learning (MML) and Domain Generalization (DG) have seen significant advancements as distinct disciplines, their systematic integration has not been thoroughly investigated. Existing MMDG studies are predominantly limited to action recognition tasks and suffer from a lack of standardized evaluation metrics.

To bridge this gap, we present MMDG-Bench, a comprehensive benchmark built upon two core frameworks: DG then MML (D2M) and MML then DG (M2D). This benchmark establishes unified experimental protocols for a variety of tasks, such as video-audio-flow action recognition and RGB-Depth-IR face anti-spoofing. By combining a standardized MML configuration with five distinct DG techniques under both D2M and M2D structures, we instantiate ten MMDG baselines. Our results show that these structured approaches often surpass current state-of-the-art methods, highlighting the critical need for unified benchmarking standards.

Our analysis reveals three primary insights: 1. The incorporation of DG techniques ensures consistent generalization improvements across various backbones, whereas models without DG are highly sensitive to changes in backbone architecture. 2. The choice between frameworks is dictated by inter-modal stability: D2M performs best when modal relationships remain stable across domains, while M2D demonstrates greater resilience to cross-domain relational variance. 3. Utilizing stronger backbones within our structured frameworks leads to amplified performance benefits.

MMDG-Bench offers a principled foundation and practical design guidelines to advance future research in multi-modal robustness. The code is available at https://github.com/qszhan/MMDG-Bench.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC