Supervised matching needs pairs that barely exist
Dense ground truth is close to impossible to annotate by hand, so the field renders its training data, ever more realistically.
Every generation chases appearance. The motion is sampled, never designed. Is realism actually what makes the data transfer?
Each step improves how the data looks: exact but flat 2-D warps, then rendered 3-D objects with real parallax and occlusion, then physically simulated scenes with scanned assets and path-traced light, then procedural worlds with no asset library at all. What no step chose is how the data moves. The displacement statistics of every one of these sources are an accident of its scene sampler, and that accident is what this dissertation measures and then manufactures.
Why render at all: how scarce real correspondence labels arenumbers
Dense ground truth cannot be clicked by hand: a 512 × 512 pair has 262,144 correspondences. Real benchmarks get theirs from sensors (LiDAR, hidden textures, tracked markers) or from a few keypoints per image, which is why every training-scale correspondence dataset is rendered. Pool sizes are the training sources of Chapter 3 (paper Table, supplement); benchmark sizes are the published dataset counts.