Abstract
In this paper, we propose a novel Global Norm-Aware Pooling (GNAP) block,
which reweights local features in a convolutional neural network (CNN)
adaptively according to their L2 norms and outputs a global feature vector with
a global average pooling layer. Our GNAP block is designed to give dynamic
weights to local features in different spatial positions without losing spatial
symmetry. We use a GNAP block in a face feature embedding CNN to produce
discriminative face feature vectors for pose-robust face recognition. The GNAP
block is of very cheap computational cost, but it is very powerful for
frontal-profile face recognition. Under the CFP frontal-profile protocol, the
GNAP block can not only reduce EER dramatically but also boost TPR@FPR=0.1\%
(TPR i.e. True Positive Rate, FPR i.e. False Positive Rate) substantially. Our
experiments show that the GNAP block greatly promotes pose-robust face
recognition over the base model especially at low false positive rate.
Users
Please
log in to take part in the discussion (add own reviews or comments).