MMDG-Bench: A Benchmark for Multimodal Domain Generalization
Researchers have introduced MMDG-Bench, a new benchmark designed to advance multimodal domain generalization. This benchmark addresses the lack of standardized evaluation in multimodal learning and domain generalization, particularly outside of action recognition. MMDG-Bench includes two frameworks, D2M and M2D, and offers unified experimental protocols for tasks like video-audio-flow action recognition and RGB-Depth-IR face anti-spoofing. The analysis reveals that integrating domain generalization techniques consistently improves generalization, and the choice between D2M and M2D depends on inter-modal stability. AI
IMPACT Provides a standardized framework and actionable guidelines for future research in multimodal robustness.