Align and Adapt: Multimodal Multiview Human Activity Recognition under Arbitrary View Combinations
AliAd integrates contrastive learning and a mixture-of-experts module to handle human activity recognition with any number of available views, including missing ones. It uses an adjusted center contrastive loss to align views based on quality, reducing complexity from O(V^2) to O(V). This allows the model to flexibly fuse diverse modalities without needing to reconstruct absent data.