For the first time we show that real age estimation can be decomposed into an apparent age estimation and an apparent to real age residual estimation, leading to improved accuracies over a standard real age estimation approach.
知识点扩充
- early fusion & late fusion
early fusion就是在特征上(feature-level)进行融合,进行不同特征的连接(concatenate),输入到一个模型中进行训练;late fusion指的是在预测分数(score-level)上进行融合,做法就是训练多个模型,每个模型都会有一个预测评分,我们对所有模型的结果进行fusion,得到最后的预测结果。常见的late fusion方法有取分数的平均值(average)、最大值(maximum)、加权平均(weighted average),另外还有采用Logistics Regression的方法进行late fusion。总之,方法很多,可视情况采取。
- two contests
Chalearn LAP competition
revealed the real power of deep learning for age estimation (at least for the feature extraction part)。
CVPR 2016 competition, the top ranked participants used a VGG-16 [28] pre-trained model。 - dataset
- It is also interesting to note that the apparent age is on average larger than real age for young adults but lower for the elderly. This is consistent with the effort made by young adults to appear more mature while the elderly attempt to look younger.