kjanjua26/git-loss-for-deep-face-recognition

This repository contains code for my paper "Git Loss for Deep Face Recognition".

Git Loss For Deep Face Recognition

---

This repository contains code for my paper "Git Loss for Deep Face Recognition".
The preprint of the paper can be found at: HERE

Usage

Standard Gitloss methodology:

Following dependencies are required for the code:

1. Tensorflow >= 1.4
2. Numpy
3. Matplotlib

This code is for MNIST (mentioned in paper). For VGGFace2, you may use the code provided by facenet. Use the function `get_git_loss()` from the `gitloss.py` in facenet to use VGGFace2.

To run the code: python3 gitloss.py

Results

Following are the plots of Gitloss and Centerloss [1].

Gitloss:

Centerloss:

VGGFace2

TODO

Citation

If you use the loss function described, please cite the paper. Please note that this is the preprint version, the published version will be out soon.

@article{calefati2018git,
  title={Git Loss for Deep Face Recognition},
  author={Calefati, Alessandro and Janjua, Muhammad Kamran and Nawaz, Shah and Gallo, Ignazio},
  journal={arXiv preprint arXiv:1807.08512},
  year={2018}
}

Contact

If there is any problem with the code or if you may have any question, feel free to open an issue or reach out here: [email protected]

References

[1] https://ydwen.github.io/papers/WenECCV16.pdf

git-loss-for-deep-face-recognition's Issues

λ G L G, the loss2(maximizing the inter-class distances) could not decline?

The loss2 always floating in a small range? how can solve it?

@Shahnawazgrewal @kjanjua26

some questiones about Git loss ( λ G )

@kjanjua26 @Shahnawazgrewal

have you find some good ways to solve the long tail ?

how about subsitute embeddings for the prelogits in the network( logits = .........)

Identities of VGGFACE2 overlapping with LFW?

Thank you for the work. Can you share with us the list for removing identities of VGGFACE2 overlapping with LFW?

Why do you call it "Git"?

Hello,

Great job on implementation and the paper. Why do you call it "Git" is it short for something in particular?

Is there a pytorch version of git loss?

Is there a pytorch version of git loss?When I use gitloss for scene classification, the gitloss value does not decline, I don't know if it is a problem I implemented or gitloss itself is not suitable for scene classification.