FV-COR: Full-View Consistency Constrained Multi-View Representation Learning for Enhanced Recognition Performance

Multi-view representation learning plays a critical role in tasks requiring robust and discriminative feature extraction across multiple perspectives. However, in real-world scenarios, multi-view data often present significant challenges, including varying viewpoint coverage, partial occlusions, inter-view discrepancies, and heterogeneous sensor modalities. Existing methods often neglect the comprehensive consistency across all views and the direct impact of learned representations on downstream recognition performance. In this paper, we propose FV-COR, a novel framework that enforces full-view consistency constraints to align both global and local features across multiple views. Furthermore, we introduce a dynamic view weighting mechanism to adaptively fuse information from heterogeneous perspectives and integrate a differentiable recognition performance optimization to directly enhance downstream tasks such as classification and retrieval. Extensive experiments on benchmark multi-view datasets demonstrate that FV-COR significantly improves recognition accuracy and representation quality compared to state-of-the-art approaches.