Mean-Variance Loss for Deep Age Estimation from a Face
Hongyu Pan, Hu Han, Shiguang Shan and Xilin Chen
CVPR 2018
None
age regression or classification do not utilize the distrubutions robustness or ordinal information in age
Use 3 loss
- original cross_entropy loss
- mean loss, softmax output * range_j
- variance loss, shrinkt probabilities
Test in Morph II and FGNet
- softmax expectation from deep expection penalize ages differently
- why use alexnet and vgg in 2018?