Reproduce “All the attention you need: Global-local, spatial-channel attention for image retrieval” 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

Chull Hwan Song, Hye Joo Han, Yannis Avrithis

Researchers at Odd Concepts and Inria developed the Global-Local Attention Module (GLAM), a unified architecture that concurrently employs global and local, spatial and channel attention for instance-level image retrieval. This framework achieved improved state-of-the-art performance on standard benchmarks, for instance, reaching 78.6 mAP on ROxford (Medium), an improvement over the 72.8 mAP baseline.

Key insights

  • Data quality is paramount for training, with cleaner datasets like GLDv2-clean yielding superior performance compared to larger, noisier alternatives.
  • While individual attention components can have varying impacts, a synergistic interaction exists where local attention, when combined with global attention, contributes significantly to overall performance.
  • Effective design choices, such as multi-scale dilated convolutions for local spatial attention, group-size batch sampling, and multi-resolution inference, are critical for achieving optimal results.

Summary by alphaXiv.

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