SLIDE 36 OF 37
What this dissertation established
Chapter 2
Data is the limiting factor in cross-modal aerial perception
Three challenge editions, one fixed dataset family: scores spread where aligned pairs are plentiful and collapse into one band where they are scarce. Our token translation improves the scarce-data directions while leaving a substantial gap.
Chapter 3
Motion coverage predicts correspondence transfer
A training-free directed distance ranks sources at 92% of the observed retraining agreement (0.6 PCK regret), extends to absolute prediction, and steers two unrelated generators as a design objective.
Chapter 4
Missing motion can be diagnosed and manufactured
An evaluation-blind doctor–pharmacist loop lifts geometric accuracy on 13/13 training sets and semantic on 12/13; a prescribed arm is best or within noise in 98% of comparisons.
As rendering grows cheap, what to render — and how it should move — matters more than how it looks.
Figures are the deck's own; the interactive instruments run in your browser from embedded data. Motion Sketch runs the pharmacist and generator as a live experiment in this companion.