Face recognition is an interesting task for computer vision apps because it gives the face of a person that you want to analyse. Also, a system like this can be combined with FaceX system in order to get more informations about the face of a person.
The purpose of this repository is to show you the performance of FaceNet on LFW dataset. In order to see the whole code you have to check the LFW & FaceNet - performance analysis.ipynb notebook. In it you have all you need to know for getting insights about the performance of LFW dataset on Face Recognition task.
The directory lfw contains the images of LFW Dataset, but not all images because on GitHub you cannot push more than 100MB of files. Also, in the facenet-weights.h5 you can find the weights of FaceNet.
This approach of using FaceNet as embedder is very nice, because it offers you the posibillity of not training again the network when a new face comes into play. Also, after my analysis I realised that how important is the choice of reference image. Reference image is that image that you want its embeddings and see the difference between others. The difference can be calculated as the euclidean/cosine/l2_euclidean distance between two faces.
There are 4 steps for face recognition task to realise:
- Face detection
- Face alignment
- Face representation (FaceNet)
- Face verification (distance)
In the below table we can see the performance of FaceNet with first image of each person as reference image. You can see in the notebook how the performance on different reference images is. Also, I want to mention that the face alignment step wasn't made in this analysis.
- TensorFlow - version 2.2.0
- OpenCV - version 4.1.2
- matplotlib - version 3.2.1
- seaborn - version 0.10.1
- pandas - version 1.0.4
- numpy - version 1.18.4
- dlib - version 19.18.0
- virtualenv - version 20.0.16
The setup is simple, you need:
- Create a virtual environment
python -m venv name_of_your_envor You can run in Google Colab the LFW & FaceNet - performance analysis.ipynb notebook. - Install the above packages with
pip install package_name
After you install the necessary packages you can run the app like this: If you choose to run on your laptop or PC, I encourage you to install all necessary packages and just after that try to run the notebook.
- Improve the performance of the model when the face is occluded or any other situations.
- Add face alignment step in order to get more accurate results.
If you want to add any improvements feel free to reach me at paul_damsa9@yahoo.com.