MAND: Modality-Aware Novelty Detection for Open-World Egocentric Activity Recognition
Researchers have developed MAND, a novel framework for open-world egocentric activity recognition that improves the detection of new activities by better utilizing multimodal data. The system addresses limitations in existing methods where visual data often overshadows other modalities like IMU, leading to underutilization of complementary evidence. MAND employs Modality-aware Adaptive Scoring (MoAS) for refined novelty detection and Modality-aware Representation Stabilization Training (MoRST) to maintain modality distinctiveness during learning, showing improved accuracy and reduced false positive rates in experiments. AI