Triplet_similarity_embedding_for_face_verification.org

Paper Information

Title

Triplet similarity embedding for face verification

Author

Swami Sankaranarayanan, Azadeh Alavi, Rama Chellappa

Published

Arxiv

Code

None

Paper Content

Why

The training time of a previous method(facenet? 1000hours) is too long. Facenet uses a triplet distance loss function and was trained on a large private dataset. This work is more general.

What

1. propose a deep network architecture and a training scheme that ensures **faster** training time.
2. **formulate** a triplet similarity embedding learning method
3. performance tested on the IJB-A datasets

How

1. Mirroring the AlexNet architecture(fewer parameters in fc layers, and using PReLU instead of ReLU)
2. Using AlexNet weights to initialize the network(see thoughts 2)
3. Using AlexNet-like network to do feature extraction.
4. Using the feature extracted(512 dims) as inputs to learn triplet similarity embedding(128 dims).
5. update a matrix $W$ to transform an intermediate 512 dims feature vector to 128 dims.

Results

The author calls the method used in facenet TDE(Triplet distance embedding), and their own TSE(Triplet similarity embedding). TSE+L2 is better than L2 only and TDE+L2

Thoughts

1. Compare running time with facenet is unfair.
2. Using VGG/ResNet/DenseNet/Inception instead of AlexNet?
3. Pay attention to the “distance” and “similarity”