/AMSoftmax

A simple yet effective loss function for face verification.

Primary LanguageMATLABMIT LicenseMIT

Additive Margin Softmax for Face Verification

by Feng Wang, Weiyang Liu, Haijun Liu, Jian Cheng

The paper is available as a technical report at arXiv.

Introduction

FeatureVis

In this work, we design a new loss function which merges the merits of both NormFace and SphereFace. It is much easier to understand and train, and outperforms the previous state-of-the-art loss function (SphereFace) by 2-5% on MegaFace.

Citation

If you find AM-Softmax useful in your research, please consider to cite:

@article{Wang_2018_amsoftmax,
  title = {Additive Margin Softmax for Face Verification},
  author = {Wang, Feng and Liu, Weiyang and Liu, Haijun and Cheng, Jian},
  journal = {arXiv preprint arXiv:1801.05599},
  year = {2018}
}

Training

Requirements: My Caffe version https://github.com/happynear/caffe-windows. This version can also be compiled in Linux.

The prototxt file is in ./prototxt. The batch size is set to 256. If your GPU's memory is not sufficient enough, you may set iter_size: 2 in face_solver.prototxt and batch_size: 128 in face_train_test.prototxt.

The dataset used for training is CASIA-Webface. We removed 59 identities that are duplicated with LFW (17) and MegaFace Set 1 (42). This is why the final inner-product layer's output is 10516. The list of the duplicated identities can be found in https://github.com/happynear/FaceDatasets.

All other settings are the same with SphereFace. Please refer to the details in SphereFace's repository.

PS: If you want to try the margin scheme described in ArcFace, you may try to transplant this layer in the experiment branch of my Caffe repository. LabelSpecificHardMarginForward() is the kernel function for cos(theta+m).

Model and Training Log

Feature normalized, s=30, m=0.35: OneDrive, Baidu Yun .

Results

See our arXiv technical report.

3rd-Party Re-implementation