Reproduce “Learning Generalisable Omni-Scale Representations for Person Re-Identification” 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

Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, Tao Xiang

Researchers from Nanyang Technological University, University of Surrey, and Queen Mary University of London introduce OSNet, a lightweight CNN architecture designed for learning omni-scale features in person re-identification. They also present OSNet-AIN, which leverages instance normalization integrated via differentiable architecture search to achieve strong domain generalization, often matching or outperforming state-of-the-art unsupervised domain adaptation methods without requiring target domain data.

Key insights

  • Effective re-ID requires an 'omni-scale' approach to feature learning, dynamically combining information from different spatial scales rather than relying on fixed or homogeneous multi-scale streams.
  • Strategic placement of Instance Normalisation layers, identified through differentiable architecture search, is crucial for mitigating domain gaps and achieving adaptation-free domain generalization in re-ID.
  • A lightweight architecture can achieve state-of-the-art re-ID performance with significantly fewer parameters (e.g., 2.2M parameters for OSNet), making it practical for real-world, resource-constrained deployments.

Summary by alphaXiv.

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