Motion coverage predicts transfer; appearance does not
Within-context Spearman ρ between peak PCK and each negated distance. Regret = median PCK given up against an oracle best-source pick.
| Distance | Appearance (DINOv3) | Motion (bag-of-flow-vectors) | ||||||
|---|---|---|---|---|---|---|---|---|
| ρ real | ρ sem. | ρ all | regret ↓ | ρ real | ρ sem. | ρ all | regret ↓ | |
| Coverage d(T→S) | −0.19 | −0.35 | −0.26 | 11.5 | +0.60 | +0.53 | +0.57 | 0.6 |
| Off-target d(S→T) | −0.16 | −0.37 | −0.24 | 12.3 | +0.09 | −0.07 | +0.03 | 7.6 |
| Chamfer | −0.19 | −0.34 | −0.25 | 12.3 | +0.37 | +0.43 | +0.40 | 3.0 |
| Sliced W₂ | −0.15 | −0.37 | −0.24 | 12.3 | +0.35 | +0.43 | +0.39 | 1.6 |
| FID | −0.21 | −0.35 | −0.27 | 12.3 | +0.38 | +0.43 | +0.40 | 1.6 |
| Random source pick | ≈0 | ≈0 | ≈0 | 15.1 | ≈0 | ≈0 | ≈0 | 15.1 |
Green = ours, bold = best in column. Motion coverage is the strongest predictor in the pooled table; appearance coverage is negatively correlated with transfer. Jackknife over sources and families keeps ρ in [0.43, 0.54]; coverage is not a proxy for dataset size (size correlates at −0.36).
