multi-crop training

**Multi-Crop Training** is a **data augmentation strategy in self-supervised learning where multiple crops of different sizes are extracted from each image** — typically 2 large global crops (covering 50-100% of the image) and several small local crops (covering 5-20%), both processing through the network. **How Does Multi-Crop Work?** - **Global Crops (2)**: 224×224, covering most of the image. Processed by both student and teacher networks. - **Local Crops (6-8)**: 96×96, small patches. Processed only by the student network. - **Training Signal**: Student must match teacher's representation of global crops using both local and global crops. - **Introduced By**: SwAV, later adopted by DINO and DINOv2. **Why It Matters** - **Local-Global Correspondence**: Forces the model to learn that local patches contain information about the whole image. - **Efficiency**: Small crops are cheap to process, adding many training signals with little compute overhead. - **Performance**: Multi-crop consistently provides 1-2% accuracy improvement over standard 2-crop training. **Multi-Crop Training** is **seeing the forest from the trees** — training models to understand global image semantics from small local patches.