This repo contains the code for my bachelor thesis on face detection and recognition. It explores the ways of speeding up deep face recognition systems by switching from the traditional pipeline - detection -> alignment -> feature extraction -> classification - to oneshot regression-classification framework.
Python >= 3.6numpyPyTorch >= 1.0dlib
- Download FEI Face Database, Caltech Faces 1999 dataset, Georgia Tech face database. Organize your folders like this:
faces
├── fei
│ └── ...
├── caltech_faces
│ └── ...
└── gt_db
└── ...
-
Run
preprocess.py. After some time it will generate labels (bbox coordinates and class labels) for your data. -
Run
organize_data.py. It will downscale images (to 320px by default) and split them into trainset and testset.
Alterantively, download already preprocessed dataset that I used. It also contains images from google for 9 more subjects.
Run train.py to train the network.
Network architecture is based on YOLO. DarkNet backbone is replaced with much more popular ResNet. Bbox attributes are separated from class probabilities and, furthermore, an additional layer for face embeddings is added. This will help to implement things like ArcFace or Triplet loss in the future.
The training process is much similar to YOLO training. Loss function is constructed similatly.
The difference is that I use smooth L1 loss for localization loss, binary cross-entropy for confidence loss and cross-entropy for classification loss.
This network was trained (on Google Colab) for 25 epochs with batch size = 128.
It was compared against traditional dlib + ResNet34 pipeline.
| dlib + ResNet34 | Described net | |
|---|---|---|
| Accuracy | 97.54% | 98.15% |
| Mean IoU | 79.82% | 78.88% |
| FPS | 6.5 | 34.5 |
This numbers were obtained on laptop GTX 1050 GPU, so fps can vary. Also, I don't believe dlib uses GPU acceleration, so comparison against pipelines that use MTCNN will be less impressive.