Slide 29 experiment design
The evidence: hold the treatment fixed, vary only content
Each arm adds the same number of pairs to the same base and trains the same matcher, three seeds. Arms differ only in where the added pairs came from.
Own Real
more of the base's own source
FlyT Random
random FlyingThings-40k pairs
FlyT Selected
same diagnosis, routed to retrieval
Random SDF
same generator, diagnosis withheld
Prescribed
rigid pairs rendered to the diagnosis
Pre. + Deform
prescription with deformation on half
13 training sets spanning semantic keypoint, optical flow, driving, and point-tracking data · scored on geometric (6 flow/tracking families) and semantic (4 keypoint benchmarks) accuracy at PCK@3% · matcher CATs++, replicated on GLU-Net.