VICAL: Vicinal Consistency Alignment for Long-Tailed Visual Recognition

The `VICAL framework proposes a new long-tailed visual recognition method based on neighborhood consistency alignment. Its core mechanism is variance reduction rather than diversity maximization. The method consists of two key components: self-consistency learning and deep integrated distillation. The former suppresses unstable high-frequency information to smooth the loss surface and mitigate overfitting; the latter uses low-resolution views to promote cross-expert low-frequency semantic consensus to avoid optimization conflicts. Extensive experiments on CIFAR-LT, ImageNet-LT, and iNaturalist 2018 datasets show that VICAL performs better than existing state-of-the-art methods.