Face Recognition Person Search
Zero-shot face recognition with pretrained dlib ResNet-50 (VGGFace2) embeddings. Searches 13,233 LFW images via 128-d Euclidean distance. 18/19 correct matches at tolerance 0.55. No training required.
LFW: 13,233 images, 5,749 identities (deep-funneled)
Pretrained VGGFace2 embeddings + Euclidean distance search — zero-shot
Zero-shot face recognition — pretrained dlib ResNet-50 (VGGFace2) with Euclidean distance search.
Dataset
- ▸LFW (Labeled Faces in the Wild): 13,233 images, 5,749 identities
- ▸Deep-funneled pre-aligned variant
- ▸Query selection: identities with ≥10 photos (random seed 42)
3-Step Pipeline
- ▸Auto-select query person (≥10 photos available)
- ▸Extract 128-d ResNet-50 embedding for anchor image
- ▸Search entire person folder: Euclidean distance → threshold @ 0.55
Results (Query: Abdullah_Gul)
| Metric | Value |
|---|---|
| Photos available | 19 |
| Matched correctly | 18 / 19 (94.7% recall) |
| Distance range | 0.000 → 0.493 |
| Threshold | 0.55 |
| Inference mode | HOG (CPU, fast) |
Why Zero-Shot Works ResNet-50 trained on VGGFace2 (3M+ face images, 9K identities) learns a face embedding space where the same person clusters tightly regardless of lighting, pose, or aging. The metric learning objective ensures inter-identity distances are larger than intra-identity distances — even for identities never seen during training.
Enhancement (Optional) Average embeddings from multiple query photos → more robust anchor representation. Reduces sensitivity to a single pose/lighting condition.