3D Face Recognition on Low-Quality Data via Dual Contrastive Learning

3D face recognition has recently gained substantial attention. While many deep learning-based techniques have achieved impressive results with high-quality datasets, recognizing faces from low-quality data, often characterized by varying poses, occlusions, and temporal changes, remains a challenge, especially when captured with low-cost sensors. In this paper, we propose a novel dual contrastive learning network for 3D face recognition on low-quality data. In particular, our approach involves two sets of contrastive learning encoders, one for point cloud pairs and another for depth map pairs, and designs an attention-based feature fusion module to assign weights to the two modalities, enhancing the discriminative power of important features. In addition, we propose a joint loss function that combines the contrastive loss with the cross-entropy loss to improve the recognition rate. Comprehensive experiments demonstrate that this method achieves state-of-the-art performance across different settings.