Key Person Agglomerative Clustering

Unsupervised Identity Discovery

During a 3,000-frame shoot, a photographer captures hundreds of individuals across varying lighting, angles, and attire. BestShots automatically groups these appearances into coherent person profiles using greedy agglomerative clustering.

---

Cosine Distance Metric & Thresholds

Clustering compares 512-dimensional vector pairs using cosine distance:

Distance(A, B) = 1 - (A · B) / (||A|| * ||B||)

• Identity Threshold: Vectors with cosine distance < 0.58 are clustered into the same person identity.
• Outlier Rejection: Distances >= 0.58 spawn separate individual candidate groups.
• Centroid Recomputation: As additional frames are assigned, the identity centroid updates dynamically to accommodate slight profile rotations and lighting shifts.

---

Avatar Crop Asset Generation

For each identified KeyPerson, the pipeline selects the highest-scoring face detection as the reference avatar:

1. 25% Padding Margin: Adds a 25% safety margin around detected bounding box coordinates to ensure full hairline, chin, and ears are captured.
2. Aspect Ratio Center Compensation: Adjusts crop dimensions so circular avatars display natural centering without clipping forehead or necklines.
3. Sharp WebP Pipeline: Renders a dedicated 256×256 WebP avatar stored in Azure Blob Storage and linked to the project sidebar for 1-click subject filtering.