Jointly learning distribution and expectation in a unified framework for facial age and attractiveness estimation

Label distribution learning achieved promising results on ordinal regression tasks such as facial age and attractiveness estimation, especially using deep label distribution learning (DLDL) methods, introducing the label distribution learning into deep convolutional neural networks. However, existing DLDL methods have an inconsistency between the training objectives and the evaluation metric, so they may be suboptimal. In addition, these methods always adopt image classification or face recognition models with a large amount of parameters, which carry expensive computation cost and storage overhead. In this paper, we firstly analyze the essential relationship between two state-of-the-art methods (ranking CNN and DLDL) and show that the ranking method is in fact learning label distribution implicitly. This result thus firstly unifies two existing popular state-of-the-art methods into the DLDL framework. Second, in order to alleviate the inconsistency and reduce resource consumption, we design a lightweight network architecture and propose a unified framework which can jointly learn label distribution and regress expectation value. The effectiveness of our approach has been demonstrated on typical ordinal regression tasks including facial age and attractiveness estimation. Our method achieves new state-of-the-art results using the single model with 36x fewer parameters and 3x faster inference speed.

Automated summary

This paper investigates the use of label distribution learning in ordinal regression tasks such as facial age and attractiveness estimation, particularly through deep label distribution learning (DLDL) methods integrated into deep convolutional neural networks. Existing DLDL methods suffer from inconsistency between training objectives and evaluation metrics, resulting in suboptimal performance. Additionally, these methods tend to employ image classification or face recognition models with a high number of parameters, leading to expensive computation cost and storage overhead. This paper firstly analyzes the relationship between two state-of-the-art methods - ranking CNN and DLDL - and demonstrates that the ranking method inherently learns label distribution. It unifies these two popular methods within the DLDL framework. Furthermore, a lightweight network architecture and a unified framework are proposed to address the inconsistency issue and reduce resource consumption. The effectiveness of this approach is demonstrated on facial age and attractiveness estimation tasks, achieving state-of-the-art results with 36 times fewer parameters and 3 times faster inference speed.