Learning discriminative representations for multi-label image recognition

Mohammed Hassanin, Ibrahim Radwan, Salman H. Khan, Murat Tahtali

Abstract: Highlights•Disentangling classes’ features in the latent space from many-to-many relationship into one-to-one.•Introducing a multi-label contrastive loss function.•Experiments show the significance of MCL over the baselines.•Ablation studies demonstrate that MCL is helpful for unbalanced datasets.