I want to implement a real-time system for image comparison (e.g. compare a face with a reference one) on an Odroid. I would like to know what are the most suitable architectures for this task. I started with methods based on triplet loss (like Facenet) but I realized that a real-time solution is not feasible. Are there good, light alternatives?
The problem might not be caused by your loss function. Deep learning models tend to be computationally demanding. Mobile devices are not usually prepared for handling models with high throughput. That being said, you might try to:
For face ID, Apple is using Siamese Network. You may get a better idea here https://towardsdatascience.com/one-shot-learning-face-recognition-using-siamese-neural-network-a13dcf739e