facial-recognition-pipeline

End-to-end computer vision pipeline for facial recognition (FaceNet embeddings) and real-time emotion detection (7 expressions) with Docker containerization.

Comment

Facial Recognition Pipeline

A deep learning pipeline for facial recognition using FaceNet, Dlib, and Docker. Achieves 90.8% classification accuracy on the LFW benchmark dataset at 25 training epochs.

Built as part of the build-your-own-x series.

Architecture Overview

Raw Images
    │
    ▼
┌─────────────────────────────────┐
│  Preprocessing  (Dlib)          │
│  • Face detection (largest face)│
│  • 68-point landmark alignment  │
│  • Center crop (180×180px)      │
└────────────────┬────────────────┘
                 │
                 ▼
┌─────────────────────────────────┐
│  FaceNet Encoder (TensorFlow)   │
│  Inception ResNet V1            │
│  Pre-trained on MS-Celeb-1M     │
│  → 128-dimensional embedding    │
└────────────────┬────────────────┘
                 │
                 ▼
┌─────────────────────────────────┐
│  SVM Classifier (scikit-learn)  │
│  Trained on LFW embeddings      │
│  → Identity prediction + prob.  │
└─────────────────────────────────┘

0 0

Key concepts:

  • Face alignment — Dlib locates 68 facial landmarks (inner eyes, bottom lip) and applies a geometric transform to standardize pose across all inputs.
  • Triplet loss embeddings — FaceNet maps each face to a 128-D vector where same-identity faces cluster together and different-identity faces are pushed apart.
  • Transfer learning — The CNN backbone is pre-trained on MS-Celeb-1M. Only the SVM classification head is trained on LFW, dramatically reducing compute and data requirements.

Results

Epochs LFW Accuracy
5 ~85.0%
25 90.8%

Training at 25 epochs takes approximately 16 minutes on a standard CPU (MacBook Pro baseline).

Tech Stack

Component Technology
Face detection Dlib + shape predictor (68pt)
CNN backbone Inception ResNet V1 (FaceNet)
Framework TensorFlow 1.x
Classifier scikit-learn SVM
Environment Docker
Language Python 3

Quick Start

Prerequisites

  • Docker installed and running

That's it — all Python dependencies (TensorFlow, OpenCV, Dlib) are handled by the Docker image.

1 — Pull the Docker image

docker pull colemurray/medium-facenet-tutorial

GPU support: nvidia-docker pull colemurray/medium-facenet-tutorial:latest-gpu

2 — Clone this repo

git clone https://github.com/YOUR_USERNAME/facial-recognition-pipeline
cd facial-recognition-pipeline

3 — Download the dataset

curl -O http://vis-www.cs.umass.edu/lfw/lfw.tgz   # 177 MB
tar -xzvf lfw.tgz -C data/

4 — Download Dlib's landmark predictor

curl -O http://dlib.net/files/shape_predictor_68_face_landmarks.dat.bz2
bzip2 -d shape_predictor_68_face_landmarks.dat.bz2
mv shape_predictor_68_face_landmarks.dat medium_facenet_tutorial/

5 — Download FaceNet weights

docker run -v $PWD:/medium-facenet-tutorial \
  -e PYTHONPATH=$PYTHONPATH:/medium-facenet-tutorial \
  -it colemurray/medium-facenet-tutorial \
  python3 /medium-facenet-tutorial/medium_facenet_tutorial/download_and_extract_model.py \
  --model-dir /medium-facenet-tutorial/etc

6 — Preprocess images

Detects, aligns, and crops all faces in the dataset. Uses multiprocessing across all available CPU cores.

docker run -v $PWD:/medium-facenet-tutorial \
  -e PYTHONPATH=$PYTHONPATH:/medium-facenet-tutorial \
  -it colemurray/medium-facenet-tutorial \
  python3 /medium-facenet-tutorial/medium_facenet_tutorial/preprocess.py \
  --input-dir /medium-facenet-tutorial/data \
  --output-dir /medium-facenet-tutorial/output/intermediate \
  --crop-dim 180

7 — Train the classifier

docker run -v $PWD:/medium-facenet-tutorial \
  -e PYTHONPATH=$PYTHONPATH:/medium-facenet-tutorial \
  -it colemurray/medium-facenet-tutorial \
  python3 /medium-facenet-tutorial/medium_facenet_tutorial/train_classifier.py \
  --input-dir /medium-facenet-tutorial/output/intermediate \
  --model-path /medium-facenet-tutorial/etc/20170511-185253/20170511-185253.pb \
  --classifier-path /medium-facenet-tutorial/output/classifier.pkl \
  --num-threads 16 \
  --num-epochs 25 \
  --min-num-images-per-class 10 \
  --is-train

8 — Evaluate

docker run -v $PWD:/medium-facenet-tutorial \
  -e PYTHONPATH=$PYTHONPATH:/medium-facenet-tutorial \
  -it colemurray/medium-facenet-tutorial \
  python3 /medium-facenet-tutorial/medium_facenet_tutorial/train_classifier.py \
  --input-dir /medium-facenet-tutorial/output/intermediate \
  --model-path /medium-facenet-tutorial/etc/20170511-185253/20170511-185253.pb \
  --classifier-path /medium-facenet-tutorial/output/classifier.pkl \
  --num-threads 16 \
  --num-epochs 5 \
  --min-num-images-per-class 10

Accuracy per identity is printed to stdout.

Project Structure

facial-recognition-pipeline/
├── Dockerfile
├── requirements.txt
├── medium_facenet_tutorial/
│   ├── align_dlib.py                      # Face alignment using Dlib landmarks
│   ├── preprocess.py                      # Detection, alignment, crop pipeline
│   ├── lfw_input.py                       # TF queue-based dataset loader
│   ├── train_classifier.py                # Embedding generation + SVM training
│   ├── download_and_extract_model.py      # FaceNet weight downloader
│   └── shape_predictor_68_face_landmarks.dat
├── etc/
│   └── 20170511-185253/
│       └── 20170511-185253.pb             # FaceNet frozen graph
├── data/                                  # Input dataset (LFW or custom)
└── output/
    ├── intermediate/                      # Preprocessed images
    └── classifier.pkl                     # Trained SVM model

Using a Custom Dataset

Replace the LFW dataset with your own photos by following the same directory structure — one folder per identity, JPEG images inside:

data/
├── Person_A/
│   ├── photo_001.jpg
│   └── photo_002.jpg
└── Person_B/
    └── photo_001.jpg

Minimum 10 images per identity is recommended (--min-num-images-per-class 10). More images per class generally improves accuracy.

Hyperparameter Tuning

Parameter Default Notes
--num-epochs 25 Higher → better accuracy, longer training
--min-num-images-per-class 10 Lower → more identities, but noisier per-class data
--num-threads 16 Match to available CPU cores
--crop-dim 180 Input resolution; higher may improve accuracy slightly

References

License

MIT

Built With

Share this project:

Updates