14th Asian Conference on Computer Vision (ACCV 2018)

Scale-Varying Triplet Ranking with Classification Loss for Facial Age Estimation

Woobin Im Sungeun Hong Sung-Eui Yoon Hyun S. Yang

Korea Advanced Institute of Science and Technology (KAIST)

Figure 1. Overall network framework of our method. In the bottleneck layer, we apply the adaptive triplet ranking strategy (L_T : Eq. 6) by selecting triplets and computing the scale-varying triplet ranking loss. Our final objective jointly includes both the ranking (L_T : Eq. 6) and classification (L_C: Eq. 9) losses simultaneously.
Figure 2. Embedding space visualization of a bottleneck feature of the network by T-SNE method. Input from test instances of the MORPH database.

In the field of age estimation, CNNs have been widely exploited in a variety of different approaches. One of them is simple classification. However, the classification loss, i.e. cross-entropy loss does not reflect the ordinal characteristics of age labels; it focuses on whether the predicted label is correct, but does not care about the degree of error between a prediction and its target value. To address the issue, we take a feature learning approach by an end-to-end learning objective for CNN, which is configured jointly from the proposed ranking constraint as well as the classification loss. Figure 1 shows the overall framework of our method.

By applying our method we can gather better features from face images (see figure 2), and better age estimation results.


In recent years, considerable efforts based on convolutional neural networks have been devoted to age estimation from face images. Among them, classification-based approaches have shown promising results, but there has been little investigation of age differences and ordinal age information. In this paper, we propose a ranking objective with two novel schemes jointly performed with an age classification objective to take ordinal age labels into account. We first introduce relative triplet sampling in which a set of triplets is constructed considering the relative differences in ages. This also addresses the problem of having limited triplet candidates, that occurs in conventional triplet sampling. We then propose the scale-varying ranking constraint, which decides the importance of a relative triplet and adjusts a scale of gradients accordingly. Our adaptive ranking loss with relative sampling not only lowers the generalization error but ultimately has a meaningful performance improvement over the state-of-the-art methods on two well-known benchmarks.