Classical vs learned matching under known changes (mean of 4 photos)
What one keypoint attends to in the other image, layer by layer (LightGlue)
cross-attention layer
Depth from a single photo: Depth Anything V2 (DINOv2 encoder)
photo ↔ depth