• Impoving triplet sample strategy: Assume identities list as class_indices
  • We iterate all person identities class_indices in an epoch. Then for each identity in a batch, we randomly sample its coresponding K images.
  • At the beginning of the epoch, we shuffle class_indices.
• Triplet loss can be optimized togother with softmax loss: run bash train_triplet_with_softmax.sh

• This implementation supports multi-gpus.
• The Data Sample is implemented by tf.data, which can load images fast in format of jpg, png.
• The Triplet Selection is implemented by tf.py_func and it's flexible to modify as you need.
• It's fast to train and the average time of one iteration with multi-gpus (210*4 images) is just about 0.6s in V100.
• It achieves 98.3% accurace on LFW with softmax pretrain(97.1%).

This project provide an efficient implementation for deep face recognition using Triplet Loss. When trained on CASIA-Webface and tested on on LFW,this code can achieve an 98.3% accuracy with softmax pretrain and 98.6% with CosFace pretrain. The framework using triplet loss can be seen as the following figure. It contrains Data Sample, Feature Extractor, Triplet Selection and Triplet Loss modules. The details can be seen as our technical report: MassFace: an effecient implementation using triplet loss for face recognition

Framework using triplet loss to train face recognition model: total P ∗ K images are sampled for P persons with K images each person. The sampled images are mapped into feature vectors through deep convolutional network. The indexs of triplet pairs are computed by a hard example mining process based on the feature vectors and the responding triplet feature pairs can be gathered. Finally, the features of triplet pairs are inputed into triplet loss to train the CNN.

Data Sample: we sample total P*K images for P persons with K images each in an iteration. We implement it by tensorflow tf.data api.

Feature Extractor: We use MobileFacenets [1] to extract the feature of the input image as a deep representation. It only has about 6.0MB parameters and can be infered very fast as well.

Triplet Selection: Triplet selection aims to choice the valid triplet (i, j, k) which is used as input of triplet loss. The valid triplet menas that i, j have the identity and i, k have different identity. We implement serveral mining strategy to select triplet pairs.

• Batch All
• Batch Random
• Batch Min Min
• Batch Min Max
• Batch Hardest

Triplet Loss:

Mining methods:The triplet selection is based on a pool of the feature vectors. There are serveral methods to obtain the pool of feature vectors with size B.

• Online
• Offline
• Semi-online

• Install tensorflow 1.8: pip install tensorflow==1.8.0
• Get code of this respository: git clone https://github.com/yule-li/MassFace.git

• Align webface and lfw dataset to 112x112(casia-112x112,lfw-112x112) using insightface align method
• If you want to train this code on your dataset, you just need to keep your dataset having the same directory like webface in which all images of the same identity located in one folder.

• Pretrain with softmax loss: run ./train_softmax.sh with argument NETWORK=mobilenet
• Train with triplet loss: modify the PRETRAINED_MODEL to your pretrained model with softmax loss and run ./train_triplet.sh

• Modify MODEL_DIR in test.sh to your trained model
• ./test.sh

1. MoibleFaces
2. MobileFaceNet_TF
3. CosFace