Skip to content
☰
Navigate
02 · Bridging the Aerial Sensor Gap
/
Evidence · 8
←
Slide
Choose a slide…
01 · Measuring and Manufacturing the Motion to Improve Visual Correspondence
02 · One task, many names
03 · Under the hood, matching means comparing features
04 · Supervised matching needs pairs that barely exist
05 · But the motion is not noise — it has structure
06 · The motion that a generator produces predicts transfer better than the realism of its renderings.
07 · From a data constraint to a design method
08 · Aerial sensing pushes domain shift to its extreme
09 · Aerial results expose the remaining data gap
10 · Our approach: translate in a discrete token space
11 · Cross-collection translation leaves a larger gap
12 · Where the aligned data runs out, every architecture fails alike.
13 · So we manufacture the data
14 · Cover the Motion, Not the Look
15 · Every dataset gets a motion fingerprint
16 · Direction matters: coverage, not symmetric distance
17 · The transfer study: 684 measurements
18 · Motion coverage predicts transfer; appearance does not
19 · Coverage approaches observed retraining agreement
20 · Motion moves transfer; appearance barely does
21 · Coverage as a design objective
22 · Before improving how a source renders, measure how well its motion covers the target's.
23 · From targeting a benchmark to filling training-set gaps
24 · Just What the Doctor Ordered
25 · A doctor and a pharmacist run the loop
26 · The doctor: every annotation is a displacement, binned jointly
27 · The coverage table, drawn rather than counted
28 · The pharmacist and the generator
29 · The evidence: hold the treatment fixed, vary only content
30 · Prescribed pairs win almost everywhere
31 · The controls tie the gains to the diagnosis
32 · Where each addition lands in motion space
33 · The dose was right before any model trained
34 · Not all missing motion is worth the same
35 · Diagnosis is what makes motion designable. Fill what is missing and the model improves — whatever rendered it.
36 · What this dissertation established
37 · Measure the motion a dataset holds. Manufacture the motion it is missing.
→
Home ↗
Slide 8
the setting
Aerial sensing pushes domain shift to its extreme
The same vehicles in EO (top) and SAR (bottom): almost no shared low-level appearance statistics.
← Previous
Chapter overview
Next →
The same vehicles in EO (top) and SAR (bottom): almost no shared low-level appearance statistics.