Risk-Bounded Distribution Reconstruction: Stable Statistic Calibration for Long-Tailed Recognition

Long-tailed recognition suffers from extreme class imbalance, where scarce tail data leads to biased and fragile feature distributions that exacerbate confusion with semantically or visually similar classes. Prior feature-space reconstruction methods transfer head-class structure or train conditional generators to synthesize tail features, yet the resulting \emph{statistical} updates are often heuristic and can degrade multi-class separability when tail estimates are unreliable. Given this issue, we propose \emph{Risk-Bounded Distribution Reconstruction} (RBDR), an offline statistic calibration framework for the two-stage long-tailed pipeline, grounded in an analysis of rival-induced discriminative directions. RBDR performs (i) risk-aware mean calibration by softly projecting any candidate update onto a supportive set such that a surrogate discriminative-risk upper bound does not increase, and (ii) covariance control by shrinking dispersion in a rival subspace while preserving diversity in orthogonal components. In the frozen-feature reconstruction setting, these calibration components convert heuristic reconstruction cues into controllable distributional updates, improving both performance and stability across multiple long-tailed benchmarks.

Real-world image datasets are often uneven: some categories have many examples, while rare categories have only a few. This makes recognition systems less reliable for rare categories, which can matter when uncommon cases are important. Our paper studies how to improve recognition under this kind of imbalance. After training a model once, we examine the patterns it has learned for each category and carefully refine the estimates for rare categories. The key idea is to avoid corrections that make a rare category look more like categories it is often confused with. We identify these likely “rival” categories and use them as safety checks when generating extra training examples for rare categories. Across several standard image recognition test datasets, this approach improves accuracy and stability, especially for rare categories. Because the refinement is done after the main model has been trained, the process is easier to inspect and control before the final decision layer is updated.