Semantic Segmentation on NYU v2 (mIoU and Δp)
68.2mIoUDUNE
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DUNEEvaluation Protocol=linear layer, DINO teacher projector=true2025.03 | 68.2 | — | |
| DUNE (no proj.)Evaluation Protocol=linear layer, DINO teacher projector=false2025.03 | 66.1 | — | |
| MASt3REvaluation Protocol=linear layer2025.03 | 60.2 | — | |
| Pri3DEvaluation Protocol=linear layer2025.03 | 54.8 | — | |
| Multi-Modal MAE + AtteNTModel=Multi-Modal MAE, AtteNT=w, Pretraining Time=980m2026.02 | 52.6 | — | |
| Multi-Modal MAEModel=Multi-Modal MAE, AtteNT=w/o, Pretraining Time=1234m2026.02 | 51.9 | — | |
| ETR-NLPBackbone=SegNet, #P (M)=25.12023.08 | 20.37 | 13.6 | |
| MTANBackbone=SegNet, #P (M)=44.42023.08 | 17.72 | 5.7 | |
| Max roamingBackbone=SegNet, #P (M)=25.12023.08 | 17.4 | 6 | |
| Task routingBackbone=SegNet, #P (M)=25.12023.08 | 16.54 | 2.7 | |
| GradNormBackbone=SegNet, #P (M)=25.12023.08 | 16.13 | -10.6 | |
| Atten.Backbone=SegNet, #P (M)=25.12023.08 | 16.02 | 0.5 | |
| Hard sharingBackbone=SegNet, #P (M)=25.12023.08 | 15.98 | 0 | |
| Cross-StitchBackbone=SegNet, #P (M)=75.32023.08 | 14.71 | -6 |