Reproduce “Large-Scale Long-Tailed Recognition in an Open World” 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

Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, Stella X. Yu

This research defines and addresses Open Long-Tailed Recognition (OLTR), a task integrating imbalanced classification, few-shot learning, and open-set recognition. It introduces a dynamic meta-embedding algorithm with modulated attention, achieving consistent outperformance across new large-scale benchmarks like ImageNet-LT, Places-LT, and MS1M-LT, with notable gains in few-shot and open-set scenarios.

Key insights

  • A unified approach to imbalanced classification, few-shot learning, and open-set recognition is essential for realistic visual data, rather than addressing these challenges in isolation.
  • Combining direct features with meta-learned memory features allows for effective knowledge transfer from data-rich head classes to data-poor tail classes.
  • Dynamically scaling meta-embeddings based on feature proximity to known class centroids provides a robust mechanism for distinguishing between known (head/tail) and unknown (open-set) categories.

Summary by alphaXiv.

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