Reproduce “Category Discovery: An Open-World Perspective” 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
Zhenqi He, Yuanpei Liu, Kai Han
This survey paper provides the first comprehensive review of Category Discovery (CD), an open-world learning paradigm that automatically categorizes unseen classes in unlabelled data using knowledge from labelled data. It introduces a multi-faceted taxonomy, systematically analyzes core methodological components, and offers a quantitative benchmark of existing methods.
Key insights
- Leveraging part-level information, large-scale self-supervised pre-trained backbones (e.g., DINOv2), hierarchical inductive biases, and auxiliary cross-modal information consistently improves Category Discovery performance.
- No single label assignment strategy (parametric vs. non-parametric) has been universally optimal, and accurately estimating the number of novel classes remains a significant challenge for most methods.
- Current CD research largely focuses on single-object image datasets, with significant challenges remaining for scaling to complex multi-object scenes and more realistic, heterogeneous open-world settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.