Majority-consensus and harmony-guided evidential multimodal classification

Evidential Deep Learning (EDL) has recently garnered significant attention due to its dual ability to identify com plex hidden patterns in data and quantify predictive uncertainty within a single forward pass. By interpreting the output of unimodal EDL as a multinomial opinion, several efforts have attempted to extend this framework to the multimodal setting. However, most of these works assume the reliability of all modalities and rely on Dempster's rule of combination, which is notoriously sensitive to conflict-a limitation that becomes particu larly problematic in real world scenarios, where multimodal data is often noisy, heterogeneous, and inherently conflicting. Under such conditions, Dempster's rule can produce counterintuitive or unreliable outcomes. To over come this limitation, we propose the Evidential Multi-Modal Classifier (EMMC). At its core lies a novel fusion strategy, the Majority-Consensus Fusion (mcf) rule, which combines Smets' unnormalized conjunctive rule with majority voting. This hybrid approach effectively bounds the influence of extreme or outlier opinions, thereby enhancing robustness in conflict-prone settings. Furthermore, the training process is guided by a novel Degree of Harmony dh-based regularization term, which encourages agreement among fused opinions. Extensive exper imental validation on both vector-type multi-view datasets and RGB-D image datasets confirms the effectiveness of our approach. The code is publicly available at https://github.com/houdaifabou/EMMC.