Reproduce “Balanced Meta-Softmax for Long-Tailed Visual Recognition” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jiawei Ren, Cunjun Yu, Shunan Sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, Hongsheng Li

Balanced Meta-Softmax (BALMS) introduces a framework for long-tailed visual recognition by addressing Softmax bias with a theoretically grounded loss function and adaptively re-sampling using meta-learning. The method achieved state-of-the-art performance across multiple long-tailed image classification benchmarks and significantly improved average precision on rare classes in the challenging LVIS instance segmentation dataset.

Key insights

  • Standard Softmax is inherently biased under long-tailed training distributions, which leads to suboptimal performance on balanced test sets.
  • Optimizing for a multi-class generalization error bound implies a need for larger decision margins for classes with fewer training samples (tail classes).
  • Combining Balanced Softmax with naive class-balanced sampling can paradoxically lead to "over-balance," where minority classes are over-emphasized during optimization, highlighting the need for adaptive sampling.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.