SAAN

This is an implementation of the Self-attention aggregation network as described in the SAAN paper.

SAAN is a neural network architecture that solves face template aggregation problems using self-attention mechanisms. Particularly, we employ Transformer implementation for the sequence encoding.

Contents

• Contents
• Detailed information
• Results on benchmarks

Detailed information

1. Model training and evaluation

   There are two notebooks which demonstrate single and multi-identity aggregation models with the respective training and validation pipelines created using tf.estimator and tf.Dataset APIs. The aggregation architecture is shared and could be found in aggregator.py.

   config.py specifies the configurations of the data sampler and different aggregators.

2. Results on benchmarks

   IJB notebook contains the complete report with the respective visualization on identification and verificaiton metrics using IJB-C benchmark.

Note: It is advised to download the repository and display the .html files via the browsers in order to view the results.