Region-Coarsening Image Embeddings

Progressively merge connected image regions and measure how Vision Transformer embeddings drift as pixel detail is removed.