Face Recognition Evaluation Pipeline using InsightFace 🎯

A production-ready, multi-model face recognition evaluation pipeline using SCRFD face detection, ArcFace embeddings, and pure cosine similarity matching evaluated across 3 diverse datasets.

Features

• 🔍 SCRFD face detector (loaded once across all models)
• 🧠 5 InsightFace models evaluated fairly on identical face crops
• 📐 Identity mean embeddings (gallery-level representation)
• ⚡ Vectorized cosine similarity matching (NumPy)
• 📊 Multi-dataset evaluation (ORL, IMDB-Crop, IMFDB)
• 💾 Decoupled pipeline: extract embeddings first, match second
• 🗂️ YAML config for reproducibility

---

Architecture

Detection (SCRFD) → Face Crop → Embedding (ArcFace/MobileNet) → Cosine Matching → Rank-1 Accuracy

---

Results

Model	ORL Rank-1	IMDB Rank-1	IMFDB Rank-1	Avg Accuracy
buffalo_l	1.0000	0.8395	0.9228	0.9208 ⭐
buffalo_s	1.0000	0.8202	0.8607	0.8937
buffalo_sc	1.0000	0.8202	0.8607	0.8937
buffalo_m	N/A*	N/A*	N/A*	0.0000
antelopev2	N/A*	N/A*	N/A*	0.0000

Best Model: buffalo_l (Avg Accuracy: 0.9208) — ResNet50 ArcFace trained on WebFace600K

Note: buffalo_m and antelopev2 encountered a model initialization error during this run and are excluded from scoring.

---

Repository Structure

Face-Recognition/
│
├── face_eval_pipeline/
│   ├── scripts/
│   │   ├── build_imfdb_gallery.py     # Build reference gallery from IMFDB
│   │   ├── build_imdb_wiki_gallery.py # Build reference gallery from IMDB/WIKI
│   │   ├── extract_all_features.py    # Stage 1: Extract all embeddings per model
│   │   └── evaluate_similarities.py   # Stage 2: Cosine matching + metrics table
│   │
│   ├── src/                           # Legacy evaluation pipeline modules
│   │   ├── data_loader.py
│   │   ├── embedding.py
│   │   ├── evaluator.py
│   │   ├── matcher.py
│   │   ├── metrics.py
│   │   ├── model_manager.py
│   │   └── pose_selector.py
│   │
│   └── results/
│       ├── embeddings/                # Saved .pkl embedding files (gitignored)
│       └── similarity_metrics.json    # Final evaluation output
│
├── utils/                             # Shared utilities
│   ├── detection.py                   # SCRFD loader + face selector
│   ├── embedding.py                   # Recognition model loader + extractor
│   ├── matching.py                    # Gallery builder + cosine matching
│   └── evaluation.py                  # Metrics + table printing + JSON saver
│
├── configs/
│   └── paths.yaml                     # Centralized dataset and output paths
│
├── run_pipeline.sh                    # End-to-end pipeline runner
├── requirements.txt
└── README.md

---

Installation

git clone https://github.com/avinash064/Face-Recognition.git
cd Face-Recognition

# Install dependencies (GPU support strongly recommended)
pip install -r face_eval_pipeline/requirements.txt

Requires: Python 3.8+, CUDA-compatible GPU (tested on RTX 3050 4GB)

---

Dataset Setup

Datasets must be downloaded manually and placed in the correct directories as specified in configs/paths.yaml.

Dataset	Download URL	Expected Path
AT&T (ORL)	AT&T	face_eval_pipeline/orl_faces/
IMDB-Crop	IMDB-WIKI	datasets/imdb_wiki/imdb_crop/
IMFDB	IMFDB	datasets/IMFDB FR dataset/

---

Usage

Option A: One Command (end-to-end)

bash run_pipeline.sh

Option B: Manual Step-by-Step

Step 1 — Build Reference Galleries (run once per dataset)

python face_eval_pipeline/scripts/build_imfdb_gallery.py
python face_eval_pipeline/scripts/build_imdb_wiki_gallery.py

Step 2 — Extract Embeddings

python face_eval_pipeline/scripts/extract_all_features.py

Step 3 — Evaluate

python face_eval_pipeline/scripts/evaluate_similarities.py

---

Configuration

Edit configs/paths.yaml to point to your dataset directories:

datasets:
  orl:
    gallery: "face_eval_pipeline/orl_faces/reference_gallery"
    probe:   "face_eval_pipeline/orl_faces"
  imdb:
    gallery: "datasets/imdb_wiki/reference_gallery_imdb"
    probe:   "datasets/imdb_wiki/imdb_crop"

---

Methodology

1. Gallery Construction: The top 5–10 most frontal images per identity are selected using pose scores (yaw + pitch from InsightFace).
2. Feature Extraction: SCRFD detects the largest face in each image. All 5 recognition models then extract normalized 512-D embeddings from that same crop.
3. Identity Representation: Gallery embeddings are averaged per identity and re-normalized → Identity Mean Embedding.
4. Matching: Probe embeddings are matched against the gallery using vectorized dot-product cosine similarity.
5. Evaluation: Rank-1 accuracy (highest similarity = predicted identity).

---

Future Work

• FAISS integration for large-scale retrieval
• Multi-model embedding fusion
• ROC curve generation & AUC metrics
• Threshold tuning for open-set recognition
• Top-5 accuracy support

---

License

MIT License. See LICENSE for details.

Citation

If you use this pipeline, please cite the InsightFace project:

@inproceedings{deng2019arcface,
  title={ArcFace: Additive Angular Margin Loss for Deep Face Recognition},
  author={Deng, Jiankang et al.},
  booktitle={CVPR},
  year={2019}
}