Depth Estimation on Cityscapes
0.0121Abs. Err.FairGrad
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| FairGrad2026.06 | 0.0121 | 43.7426 | — | — | |
| STCH2026.06 | 0.0123 | 42.8384 | — | — | |
| CORE-MTL2026.06 | 0.0123 | 19.6088 | — | — | |
| OursBackbone=SegNet, Parameter Count (#P.)=95.482024.06 | 0.0125 | 41.6 | -10.08 | — | |
| STL2026.03 | 0.0125 | 27.77 | — | — | |
| GradNorm2026.06 | 0.0125 | 44.552 | — | — | |
| MTAN2026.06 | 0.0126 | 45.5663 | — | — | |
| Single-taskBackbone=SegNet, Parameter Count (#P.)=190.592024.06 | 0.0128 | 29.98 | — | — | |
| Equal Weighting2026.06 | 0.0128 | 44.034 | — | — | |
| PCGrad2026.06 | 0.0128 | 44.5574 | — | — | |
| ExcessMTL2026.06 | 0.0128 | 43.3338 | — | — | |
| MGDA w/ ReconBackbone=SegNet, Parameter Count (#P.)=108.44, With Recon=true2024.06 | 0.0129 | 33.41 | -4.46 | — | |
| Nash-MTLCost=O(md)2026.03 | 0.0129 | 35.02 | — | — | |
| Single Task2026.06 | 0.0129 | 47.9603 | — | — | |
| MGDABackbone=SegNet, Parameter Count (#P.)=95.43, With Recon=false2024.06 | 0.013 | 47.09 | -16.22 | — | |
| CAGrad w/ ReconBackbone=SegNet, Parameter Count (#P.)=108.44, With Recon=true2024.06 | 0.013 | 38.27 | -7.38 | — | |
| RLW2026.06 | 0.0131 | 43.6991 | — | — | |
| Graddrop w/ ReconBackbone=SegNet, Parameter Count (#P.)=108.44, With Recon=true2024.06 | 0.0134 | 41.37 | -10.69 | — | |
| IMTL-GCost=O(md)2026.03 | 0.0135 | 38.41 | — | — | |
| ETR-NLPBackbone=SegNet, Number of Parameters (M)=22.12023.08 | 0.0136 | 29.16 | 20.8 | — | |
| GD w/ ReconBackbone=SegNet, Parameter Count (#P.)=108.44, With Recon=true2024.06 | 0.0136 | 43.18 | -12.63 | — | |
| PCGrad w/ ReconBackbone=SegNet, Parameter Count (#P.)=108.44, With Recon=true2024.06 | 0.0136 | 46.02 | -14.92 | — | |
| SDMGradCost=O(md)2026.03 | 0.0137 | 34.01 | — | — | |
| MARIGOLDCost=O(d)2026.03 | 0.0139 | 34.79 | — | — | |
| FAMO (rerun)Cost=O(d)2026.03 | 0.014 | 35.78 | — | — | |
| CAGradCost=O(md)2026.03 | 0.0141 | 37.6 | — | — | |
| Max roamingBackbone=SegNet, Number of Parameters (M)=25.12023.08 | 0.0143 | 29.38 | 17.2 | — | |
| MTANBackbone=SegNet, Number of Parameters (M)=44.42023.08 | 0.0144 | 33.63 | 10.9 | — | |
| FAMOCost=O(d)2026.03 | 0.0145 | 32.59 | — | — | |
| MOML2026.06 | 0.0145 | 52.5924 | — | — | |
| RotoGradBackbone=SegNet, Parameter Count (#P.)=103.432024.06 | 0.0147 | 82.31 | -47.81 | — | |
| MoCoCost=O(md)2026.03 | 0.0149 | 34.19 | — | — | |
| CAGradBackbone=SegNet, Parameter Count (#P.)=95.43, With Recon=false2024.06 | 0.0153 | 88.29 | -53.81 | — | |
| Cross-StitchBackbone=SegNet, Number of Parameters (M)=75.32023.08 | 0.0154 | 34.49 | 6.5 | — | |
| PCGradCost=O(md)2026.03 | 0.0154 | 42.07 | — | — | |
| Task routingBackbone=SegNet, Number of Parameters (M)=25.12023.08 | 0.0155 | 31.47 | 12.4 | — | |
| GradDropCost=O(md)2026.03 | 0.0157 | 47.54 | — | — | |
| MoDoCost=O(md)2026.03 | 0.0159 | 41.51 | — | — | |
| Atten.Backbone=SegNet, Number of Parameters (M)=25.12023.08 | 0.016 | 35.72 | 7.6 | — | |
| Cross-StitchBackbone=SegNet, Parameter Count (#P.)=190.592024.06 | 0.0162 | 116.66 | -79.04 | — | |
| Joint trainingBackbone=ResNet-18, Task Sequence=Segmentation -> Depth2026.03 | 0.0164 | 43.7236 | — | — | |
| Joint trainingBackbone=ResNet-18, Task Sequence=Depth -> Segmentation2026.03 | 0.0164 | 43.7236 | — | — | |
| GDBackbone=SegNet, Parameter Count (#P.)=95.43, With Recon=false2024.06 | 0.0166 | 116 | -79.32 | — | |
| Hard sharingBackbone=SegNet, Number of Parameters (M)=25.12023.08 | 0.017 | 43.99 | 0 | — | |
| GraddropBackbone=SegNet, Parameter Count (#P.)=95.43, With Recon=false2024.06 | 0.0173 | 115.79 | -80.48 | — | |
| RepMTL2026.06 | 0.018 | 44.323 | — | — | |
| HADBackbone=ResNet-18, Task Sequence=Depth -> Segmentation2026.03 | 0.0186 | 45.5291 | — | — | |
| HADBackbone=ResNet-18, Task Sequence=Segmentation -> Depth2026.03 | 0.0192 | 51.7254 | — | — | |
| LwFBackbone=ResNet-18, Task Sequence=Depth -> Segmentation2026.03 | 0.0192 | 47.2005 | — | — | |
| GradNormBackbone=SegNet, Number of Parameters (M)=25.12023.08 | 0.0199 | 68.13 | -23.9 | — | |
| PCGradBackbone=SegNet, Parameter Count (#P.)=95.43, With Recon=false2024.06 | 0.02 | 114.5 | -78.39 | — | |
| LwFBackbone=ResNet-18, Task Sequence=Segmentation -> Depth2026.03 | 0.0202 | 47.64 | — | — | |
| SGPBackbone=ResNet-18, Task Sequence=Segmentation -> Depth2026.03 | 0.0202 | 49.2638 | — | — | |
| Vanilla trainingBackbone=ResNet-18, Task Sequence=Segmentation -> Depth2026.03 | 0.0203 | 50.0861 | — | — | |
| EWCBackbone=ResNet-18, Task Sequence=Segmentation -> Depth2026.03 | 0.0203 | 56.8526 | — | — | |
| iCaRLBackbone=ResNet-18, Task Sequence=Segmentation -> Depth2026.03 | 0.0204 | 50.2167 | — | — | |
| DERBackbone=ResNet-18, Task Sequence=Segmentation -> Depth2026.03 | 0.0206 | 51.3117 | — | — | |
| EWCBackbone=ResNet-18, Task Sequence=Depth -> Segmentation2026.03 | 0.0217 | 51.956 | — | — | |
| iCaRLBackbone=ResNet-18, Task Sequence=Depth -> Segmentation2026.03 | 0.0221 | 58.8481 | — | — | |
| DERBackbone=ResNet-18, Task Sequence=Depth -> Segmentation2026.03 | 0.0256 | 54.3059 | — | — | |
| MGDACost=O(md)2026.03 | 0.0309 | 33.5 | — | — | |
| SPGBackbone=ResNet-18, Task Sequence=Depth -> Segmentation2026.03 | 0.0418 | 104.0395 | — | — | |
| Vanilla trainingBackbone=ResNet-18, Task Sequence=Depth -> Segmentation2026.03 | 0.0456 | 77.8347 | — | — | |
| SPGBackbone=ResNet-18, Task Sequence=Segmentation -> Depth2026.03 | 0.0484 | 94.8094 | — | — | |
| SGPBackbone=ResNet-18, Task Sequence=Depth -> Segmentation2026.03 | 0.0484 | 94.8094 | — | — | |
| SwinMTL (Swin-B)Additions=MLP Decoder2026.01 | — | — | — | 92.1 | |
| Youtu-VL (4B)Additions=None2026.01 | — | — | — | 92.7 |