Local-Global Correspondence is a learning principle in self-supervised vision where the model is trained to predict global image properties from local patches — ensuring that every part of the image encodes information about the whole, producing rich, hierarchical representations.

What Is Local-Global Correspondence?

• Principle: A small crop of an image (e.g., a cat's ear) should map to the same representation cluster as the full image (the complete cat).
• Implementation: Cross-predict between local crops and global crops in the contrastive/distillation loss.
• Methods: DINO, SwAV, and iBOT all leverage local-global correspondence.

Why It Matters

• Semantic Features: Encourages the model to learn semantic, part-aware representations rather than texture-only features.
• Dense Prediction: Improves performance on downstream dense tasks (segmentation, detection) where local features must encode broader context.
• Emergent Properties: DINO's ability to produce segmentation masks from attention maps is attributed to local-global correspondence training.

Local-Global Correspondence is the holographic principle for vision — ensuring every pixel encodes something about the whole scene.