The Challenge
A photo archive with 12M+ uncategorized portraits required identity grouping, but manual tagging was infeasible. Variations in pose, lighting, and image quality made clustering difficult.
Our Solution
Trained a face recognition model on historical photos to generate embeddings, then used hierarchical clustering with optimized distance thresholds. Incorporated manual review loops for ambiguous cases.
Technologies Used
Results & Impact
- Clustered 12M+ photos into ~800K distinct identities
- Achieved 89% precision/recall on identity grouping
- Reduced manual tagging effort by 90%