Metric-aware latent oversampling for long-tailed image recognition

Real-world visual datasets often follow long-tailed distributions, where a few head classes contain abundant samples while many tail classes are severely underrepresented. Existing solutions predominantly adopt algorithm-centric strategies, such as re-weighting or decoupled training, but these often fail to produce robust representations for tail categories. In contrast, data-centric approaches, particularly oversampling, offer a promising alternative yet remain underexplored in vision due to the high dimensionality of embeddings and the risk of distorting head-class structure. To address these challenges, we introduce a structured metric learning framework that enriches tail-class diversity through class-conditional interpolation in the latent space. The core mechanism is a quadruplet design consisting of a anchor-positive pair from one class and an anchor-negative pair from another, enabling interpolation across both intra-class and inter-class relations. Interpolation is selectively applied to minority classes and moderated by a stochastic gating variable, ensuring adaptive augmentation without excessive perturbation. A sparsity-aware strategy selects positives to broaden intra-class coverage, while proximity-based hard negative mining refines inter-class boundaries. During inference, minority-class embeddings are further augmented in latent space, yielding more balanced and discriminative feature distributions. Extensive experiments on multiple long-tailed benchmarks demonstrate that the proposed framework achieves competitive and robust performance compared with existing long-tailed recognition methods, particularly under severe imbalance conditions, highlighting the effectiveness of a structured data-centric perspective for long-tailed visual recognition.