Improved Face Verification with Simple Weighted Feature Combination

Since the appearance of deep learning, face verification (FV) has made great progress with large scale datasets, well-designed networks, new loss functions, fusion of models and metric learning methods. However, incorporating all these methods obviously takes a lot of time both at training and testing stages. In this paper, we just select training images randomly without any clean and alignment procedure. Then we propose a simple weighted average method which combines features of the last two layers with different weights on the modified VGGNet, named as CB-VGG. It is significantly reducing the complexity of time that one model can be treated as two models. LMNN is used as a post-processing procedure to improve the discrimination of the combined features. Our experiments show relatively competitive results on LFW, CFP, and CACD datasets.