SELF-CORRECTIVE AND STABILITY-AWARE SELF-DISTILLATION FOR LONG-TAILED RECOGNITION

Abstract

For image-based object recognition, long-tailed data distributions pose a fundamental challenge to model training, as visual supervision becomes highly imbalanced across object categories. Existing learning paradigms often struggle to balance performance between head and tail classes under such imbalanced supervision. In this paper, we propose a self-corrective and stability-aware self-distillation framework for long-tailed visual recognition. The proposed method explicitly distinguishes reliable predictions that should be preserved from persistent errors that require controlled correction during training, enabling more effective learning across different class-frequency regimes. By jointly considering knowledge preservation and error correction, the proposed framework prevents severe performance degradation on head classes while progressively improving recognition accuracy on tail classes. Extensive experiments on CIFAR- 10-LT, CIFAR-100-LT, Tiny ImageNet-LT and ImageNet-LT demonstrate that our approach consistently outperforms existing long-tailed recognition methods, achieving significant improvements on tail classes while maintaining strong overall classification performance.