Introduction
Deploying face recognition models to new surveillance cameras often results in a significant drop in accuracy due to "domain shift"—variations in lighting, camera angles, and environmental conditions. Traditionally, correcting this requires manually labeling faces from the new environment, an expensive and impractical task for large-scale networks. This paper introduces Dual-Triplet Metric Learning (DTML), an unsupervised domain adaptation (UDA) technique that allows models to adapt to new cameras using only a small amount of unlabeled video data.
Methodology: Dual-Triplet Metric Learning
Face recognition typically employs Siamese networks trained with a triplet loss function, which learns to minimize the distance between images of the same person while maximizing the distance between different people. To adapt to a new, unlabeled "target" domain, DTML optimizes two triplet losses simultaneously:
- Source Triplet ($L_s$): Uses labeled data from the original training environment to maintain a discriminative baseline.
- Target Triplet ($L_t$): Uses unlabeled data from the new camera.
Mutual Supervision
To generate target triplets without ground-truth labels, the authors propose a Mutual Supervision approach. A pre-trained source model acts as a "teacher," calculating the statistical distributions (mean and standard deviation) of "within-class" (same person) and "between-class" (different people) distances in the source data. These statistics define "mining windows." When unlabeled target data is processed, pairwise distances falling within these windows are pseudo-labeled as positive or negative pairs to construct the target triplet loss.
Key Insights
- Distance Distribution Alignment: For open-set problems like face recognition, it is more effective to align pairwise distance distributions rather than absolute feature representations. By using source domain confidence intervals, the model can safely mine pairs from unlabeled target data.
- Preventing Catastrophic Forgetting: Relying solely on target triplet loss ($L_t$) can lead to overfitting on noisy pseudo-labels, causing the model to collapse. Maintaining the source triplet loss ($L_s$) acts as an anchor, ensuring the underlying feature representation remains robust during adaptation.
Experimental Results
The authors evaluated DTML on both public and private datasets:
- COX Face Dataset: In a cross-camera adaptation scenario (Camera 3 to Camera 2), the baseline model achieved 87.2% accuracy. DTML improved this to 91.0%, nearing the theoretical upper bound of 95.3% achievable with full supervision.
- Real-World IP Camera: When adapting a lab-trained model to a commercial surveillance camera, the baseline True Positive Rate (at 1% FAR) was 52%. DTML adaptation increased this to 73%. By adding an extra classification layer trained on the adapted representations, accuracy reached 84%.
Conclusion and Future Work
DTML provides a practical path for deploying face recognition in dynamic environments without manual labeling. A current limitation is the requirement for some initial overlap between source and target distance distributions. Future research will explore adversarial learning to force an initial baseline alignment before applying the statistical pseudo-labeling step, as well as end-to-end training of the entire architecture.
Summary updated