🚀🚀Real-time Face Recognition in TensorLayer

[zsdonghao]

A discussion for:

• Face recognition algorithm
• Face recognition history
• Face recognition implementation using TensorLayer and TensorFlow.

Feel free to add more papers and discuss here or in the Slack channel.

Background

SphereFace : Face recognition (FP) can be categorized as face identification and face verification. The identification classifies a face to a specific identity, while the verification determines whether a pair of faces belongs to the same identity.

For closed-set protocol, all testing identities are predefined in training set. Therefore, closed- set FR can be well addressed as a classification problem.

For open-set protocol, the testing identities are usually not in the training set, so we need to map faces to a discriminative feature space. Then face identification can be viewed as performing face verification between the probe face and every identity in the gallery (given some faces of the identities). <--- industry usually use this one.

Paper History

• [Dimensionality reduction by learning an invariant mapping. In CVPR, 2006.]()
  • triplet loss
• [Deep learning face representation from predicting 10,000 classes. In CVPR, 2014.]()
  • softmax loss, treats open-set FR as a multi-class classification problem
  • open-set
• [Deepface: Closing the gap to human-level performance in face verifica- tion. In CVPR, 2014.]()
  • softmax loss, treats open-set FR as a multi-class classification problem
  • open-set
• [Deep learning face representation by joint identification-verification. In NIPS, 2014]()
  • softmax loss + contrastive loss (Euclidean margin based loss)
  • greatly boosting the performance.
• [Facenet: A unified embedding for face recognition and clustering. In CVPR, 2015]()
  • triplet loss
  • code davidsandberg
  • learn a unified face embedding, 200 million face images, current state-of-the-art FR accuracy
• [A discriminative feature learning approach for deep face recognition. In ECCV, 2016]()
  • softmax loss + centre loss (Euclidean margin based loss)

SphereFace : One could notice that state-of-the-art FR meth- ods usually adopt ideas (e.g. contrastive loss, triplet loss) from metric learning, showing open-set FR could be well addressed by discriminative metric learning.

• [Sparsifying neural network connections for face recognition. In CVPR, 2016]()

  • softmax loss + contrastive loss (Euclidean margin based loss)

• [Targeting ultimate accuracy: Face recognition via deep embedding. arXiv preprint:1506.07310, 2015.]()

  • ? loss

• [Large-margin softmax loss for convolutional neural networks. In ICML, 2016. 2,]()

  • L-Softmax loss, also implicitly involves the concept of angles like SphereFace. Differently, SphereFace A-Softmax loss is developed to explicitly learn discriminative face embedding.
  • it shows great improvement on closed-set classification problems.

SphereFace : Center loss only explicitly encourages intra-class compactness. Both contrastive loss and triplet loss can not constrain on each individual sample, and thus require carefully designed pair/triplet mining procedure, which is both time-consuming and performance-sensitive.

• [SphereFace: Deep Hypersphere Embedding for Face Recognition. In CVPR, 2017]()
  • angular softmax (A-Softmax) loss
  • open-set
  • haijun : 100x100 still works fine
  • code wy1iu
  • We extract the deep features (SphereFace) from the output of the FC1 layer. For all experiments, the final representation of a testing face is obtained by concatenating its original face features and its horizontally flipped features. The score (metric) is computed by the cosine distance of two features.
• [CosFace: Large Margin Cosine Loss for Deep Face Recognition In ArXiv, 2018]()
• ArcFace/InsightFace: Additive Angular Margin Loss for Deep Face Recognition. ArXiv, 2018
  • jiankang : follows sphereface and cosface
  • code insightface (ArcFace) from Imperial College and DeepInsight using MXNET
  • code InsightFace_TF (ArcFace) using TensorLayer
• MobileFaceNets: Efficient CNNs for Accurate Real-time Face Verification on Mobile Devices. ArXiv, 2018

Implementation Hints

• https://github.com/auroua/InsightFace_TF/blob/master/losses/face_losses.py
• http://tensorlayer.readthedocs.io/en/latest/modules/cost.html#cosine-similarity
• https://github.com/sirius-ai/MobileFaceNet_TF

[zsdonghao]

@DEKHTIARJonathan yolo3 is for object detection

[DEKHTIARJonathan]

Oh my bad