After submission, we found that the released UniScene-v2 code sets the noise-prior weight to λ=0.05 whereas our paper reports λ=0.15. This page re-evaluates UniScene-v2 at λ=0.05. The change is marginal and does not alter any of the paper's contributions or conclusions: ROCC still outperforms UniScene-v2 by a wide margin on every metric and set. We report it for full transparency. Results at the paper's λ=0.15 are here.
| UniScene-v2 | UniScene-v2 | Dense | + bg | ROCC | ||
|---|---|---|---|---|---|---|
| Set | Metric | (λ=0.05) | (λ=0) | CNN | head‡ | (ours) |
| mini-val | fg-mIoU ↑ | 20.19 | – | 33.55 | 45.62 | 49.25 |
| mIoUUS ↑ | 29.53 | – | 38.09 | 49.01 | 51.65 | |
| occ-IoU ↑ | 22.24 | – | 24.74 | 24.19 | 27.25 | |
| key subset | fg-mIoU ↑ | 20.06 | 0.03 | 32.91 | 44.90 | 48.45 |
| mIoUUS ↑ | 29.37 | 13.12 | 37.51 | 48.35 | 50.92 | |
| occ-IoU ↑ | 21.65 | 7.31 | 23.90 | 23.40 | 26.17 |
Occupancy generation on two views of the nuPlan-Occ validation split: mini-val (the full validation set) and key subset (one key frame per clip). One shared per-class IoU accumulator scores every column on identical ground truth. UniScene-v2's λ supplies a fraction of the ground-truth occupancy latent of the scored frame; at λ=0 it collapses. ‡ trained at a larger effective batch than the other columns.