Addressing Long-Tailed Age Estimation with Balancing Techniques

In real life, the human age distribution exhibits a pronounced long-tail characteristic, which can interfere with age estimation, leading to a decrease in the accuracy of age estimation models under this long-tail characteristic. In recent years, the long-tail characteristic of age estimation datasets has become increasingly prominent, manifesting in both real life and facial datasets as a lack of samples for the elderly and children. We refer to categories with a high number of samples as head classes and those with a low number of samples as tail classes. Traditional research methods often involve forcibly enhancing the prediction weights of tail classes, which damages the prediction weights of head classes or significantly increases the complexity of the model. In response to this issue, a new method for long-tail age estimation called Balanced Prediction Constraint (BPC) method has been proposed, which is used to rebalance long-tail age estimation by imposing prediction constraints on the original dataset. This method neither increases the complexity of the model nor affects the head classes while enhancing the prediction probabilities of the tail classes. Experimental results on the Morph II dataset achieved an Mean Absolute Error (MAE) of 2.32, which is a better result compared to methods like OR-CNN and CORAL, and it remains competitive when compared to other advanced methods.