This repository contains code for fine-tuning a face recognition model using the Inception ResNet V2 architecture with a triplet loss in TensorFlow. The model is trained to recognize faces of individuals from a dataset, and it uses a face-only dataset obtained using the MTCNN (Multi-task Cascaded Convolutional Networks) face detection system.
To reproduce the results, follow the steps below:
- Clone the repository:
git clone https://github.com/carlesoctav/face-recognition-with-triplet-loss-tensorflow.git
- Install the dependencies: (use any environment management tool of your choice)
pip install -r requirements.txt
- Prepare the dataset:
For each person whose face you want the system to recognize, create a folder with the name of the person in the data folder, and put all the images of the person in that folder. For example, if you want the system to recognize the face of "Jack", create a folder named "Jack" in the data folder, and put all the images of Jack in that folder.
- Get the face-only dataset using MTCNN:
Run the following command to extract faces from the images in the dataset using MTCNN:
python script/face_detection.py
- Train the model:
Train the model to recognize the faces of the individuals in the dataset. The pretrained Inception ResNet V2 model will be fine-tuned only on the classification blocks, using the face-only dataset obtained in the previous step. Use the following command:
python script/training_model.py
You can check all available arguments using:
python script/training_model.py --help
- Create embeddings for each person in the dataset:
Run the following command to create embeddings (feature vectors) for each person in the dataset using the trained model:
python script/embedding-space-for-inference.py
- Inference time:
Run the following command to perform inference on new images using the trained model and embedding vectors that were created in the previous step:
python script/inference.py
FaceNet: A Unified Embedding for Face Recognition and Clustering
https://github.com/pooya-mohammadi/Face
Image similarity estimation using a Siamese Network with a triplet loss