These are the results at the paper's setting, λ=0.15. After submission we found the released UniScene-v2 code uses λ=0.05 by default; those (marginally different) results are on the λ=0.05 page and are the ones we now consider primary. ROCC's margin is unchanged in substance at either setting.
| UniScene-v2 | UniScene-v2 | Dense | + bg | ROCC | ||
|---|---|---|---|---|---|---|
| Set | Metric | (λ=0.15) | (λ=0) | CNN | head‡ | (ours) |
| mini-val | fg-mIoU ↑ | 18.07 | – | 33.55 | 45.62 | 49.25 |
| mIoUUS ↑ | 25.81 | – | 38.09 | 49.01 | 51.65 | |
| occ-IoU ↑ | 16.64 | – | 24.74 | 24.19 | 27.25 | |
| key subset | fg-mIoU ↑ | 18.05 | 0.03 | 32.91 | 44.90 | 48.45 |
| mIoUUS ↑ | 25.87 | 13.12 | 37.51 | 48.35 | 50.92 | |
| occ-IoU ↑ | 16.56 | 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.