Toward a Balanced Feature Space for the Deep Imbalanced Regression

Regression with imbalanced data has been regarded as a more realistic scenario due to the difficulty of data acquisition and label annotations. However, it has not been extensively studied compared to the imbalanced classification. In imbalanced regression scenario, the classical regression approach may lead to regression bias toward high-frequency target regions. In this study, we present a novel framework for effectively handling imbalanced data in regression tasks. We introduce a density-based stochastic mask that perturbs the mini-batch distribution by assigning probabilities based on the data distribution statistics. The mask assigns a higher probability to more frequent samples on the Bernoulli distribution. Next, we employ consistency-based learning to encourage the encoder to produce similar representations for perturbed versions of the same input, drawing inspiration from modern consistency-based learning approaches. By jointly training with the two proposed learning objectives, we achieved state-of-the-art performance on AgeDB-DIR and IMDB-WIKI-DIR, which are representative imbalanced age estimation datasets. Furthermore, we evaluated the generalization performance using UTKFace. Through extensive experiments, we confirmed that our method demonstrates efficacy in dealing with imbalanced regression data. The forthcoming task involves extending the suggested approach to different uses, such as predicting the progress of diseases in medical diagnoses and estimating monocular depth in self-driving technology.