MAND: Modality-Aware Novelty Detection for Open-World Egocentric Activity Recognition

The article presents MAND, a modality-aware framework designed for multimodal egocentric activity recognition in open-world environments. Key features include Modality-aware Adaptive Scoring (MoAS) for dynamic modality contribution adjustments and Modality-aware Representation Stabilization Training (MoRST) to maintain modality-specific performance across tasks. MAND demonstrates improved novelty detection and accuracy on a public benchmark, addressing issues of catastrophic forgetting and underutilization of non-RGB modalities, which is critical for practitioners developing robust AI systems in dynamic settings.