Are We Making Progress in Multimodal Domain Generalization? A Comprehensive Benchmark Study
The study introduced MMDG-Bench, a comprehensive benchmark to test Multimodal Domain Generalization methods rigorously across six datasets and nine models. Findings show that current specialized methods offer only marginal gains over baseline, and no single method excels universally. Crucially, all tested models suffer significant performance drops when faced with real-world issues like data corruption or missing modalities.