Reproduce “Deep Learning for Cross-Domain Few-Shot Visual Recognition: A Survey” 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
Huali Xu, Shuaifeng Zhi, Shuzhou Sun, Vishal M. Patel, Li Liu
This survey provides the first comprehensive review of Cross-Domain Few-Shot Visual Recognition (CDFSL), formalizing its unique challenges and proposing a unified taxonomy for existing methods. It analyzes the empirical performance of these approaches across various benchmarks, highlighting their strengths and weaknesses in different domain transfer scenarios.
Key insights
- CDFSL faces an "unreliable two-stage empirical risk minimization" problem, stemming from limited target data and significant domain/task shifts, which renders standard empirical risk minimization ineffective.
- Existing CDFSL methods can be systematically categorized into D-Extension (augmenting information), H-Constraint (constraining hypothesis space), -Adaptation (reducing domain/task distance), or Hybrid approaches, based on how they mitigate the core problem.
- Different CDFSL strategies exhibit varying effectiveness depending on the domain distance and task type; for instance, -Adaptation performs well in distant domains, while D-Extension excels in fine-grained tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.