Delving into Deep Imbalanced Regression

Summary

Introduces the first systematic framework for deep imbalanced regression (DIR) — the problem of training neural networks on regression targets with highly skewed distributions, where rare target values are underrepresented. Proposes two complementary techniques: Label Distribution Smoothing (LDS) which re-weights training samples by convolving the empirical label distribution with a kernel, and Feature Distribution Smoothing (FDS) which calibrates feature statistics across the target space. Validated on age estimation from facial images, head pose, depth estimation, and critically, ECG-based heart rate estimation — where the HR distribution is imbalanced (common HRs over-represented).