OpenMvL: An Attribute-Inspired Dual-Head Framework for Open-Set Multiview Learning

Numerous researchers have sought to integrate multiview data to enhance task performance. However, existing multiview methods are primarily designed for closed-set environments with fully known classes. In real-world open-set scenarios, the emergence of unknown classes leads to intra-view heterogeneity and inter-view discriminative conflicts. To address these issues, we propose a novel attribute-inspired dual-head framework for open-set multiview learning. First, we develop a multiview attribute-inspired open-set network architecture: 1) A view-specific deep unfolding layer resolves intra-view conflicts by eliminating redundant features and achieving heterogeneity-reduced, consistency-based topological representations for each view; 2) An evidential fusion manner tackles inter-view conflicts by quantifying and integrating global uncertainty, selectively fusing complementary features to create a more discriminative cross-view representation. Furthermore, we design a dual-head perception-adaptive open-set training loss to further bridge these conflicts: 1) the $Softmax$ -head loss optimizes representation-related parameters and enhances the quality of representation learning, consolidating inherent open-set attributes within the representations, and 2) the $Evidence$ -head loss optimizes uncertainty-related parameters and calibrates global uncertainty, assisting to disentangle uncertainty from open-set discriminative information. Comprehensive evaluations on challenging open-set multiview tasks demonstrate the method's outstanding performance, effectively adapting to both known and unknown samples.