Depth Estimation on SUN RGB-D
0.386Depth Errorlayer specialization
Evaluation results
| Method | Links | |
|---|---|---|
| layer specializationFramework=DINOv2, Size=H, Handling Strategy=attention bias2026.02 | 0.386 | |
| DINOv2Framework=DINOv2, Size=H, Handling Strategy=attention bias2026.02 | 0.387 | |
| Ours (Specialization)Backbone=ViT-L, Framework=DINOv2, Attention Bias=false2026.02 | 0.39 | |
| layer specializationFramework=DINOv2, Size=L, Handling Strategy=none2026.02 | 0.39 | |
| Attn. bias + Ours (Specialization)Backbone=ViT-L, Framework=DINOv2, Attention Bias=true2026.02 | 0.393 | |
| layer specializationFramework=DINOv2, Size=L, Handling Strategy=attention bias2026.02 | 0.393 | |
| layer specializationFramework=DINOv2, Size=L, Handling Strategy=4 registers2026.02 | 0.395 | |
| Attn. bias + CaiTBackbone=ViT-L, Framework=DINOv2, Attention Bias=true2026.02 | 0.4 | |
| Attn. bias baselineBackbone=ViT-L, Framework=DINOv2, Attention Bias=true2026.02 | 0.406 | |
| DINOv2Framework=DINOv2, Size=L, Handling Strategy=attention bias2026.02 | 0.406 | |
| DINOv2 baseline (∅)Backbone=ViT-L, Framework=DINOv2, Attention Bias=false2026.02 | 0.41 | |
| DINOv2Framework=DINOv2, Size=L, Handling Strategy=none2026.02 | 0.41 | |
| DINOv2Framework=DINOv2, Size=L, Handling Strategy=4 registers2026.02 | 0.411 | |
| CaiTBackbone=ViT-L, Framework=DINOv2, Attention Bias=false2026.02 | 0.421 | |
| MuRF (Ours)Arch.=ViT-B/14, Evaluation Protocol=lin. 42026.03 | 0.432 | |
| High ResolutionArch.=ViT-B/14, Evaluation Protocol=lin. 42026.03 | 0.433 | |
| Medium ResolutionArch.=ViT-B/14, Evaluation Protocol=lin. 42026.03 | 0.434 | |
| iBOTArch.=ViT-L/16, Evaluation Protocol=lin. 42026.03 | 0.435 | |
| MuRF (Ours)Arch.=ViT-B/14, Evaluation Protocol=lin. 12026.03 | 0.441 | |
| iBOTArch.=ViT-L/16, Evaluation Protocol=lin. 12026.03 | 0.447 | |
| Medium ResolutionArch.=ViT-B/14, Evaluation Protocol=lin. 12026.03 | 0.453 | |
| Low ResolutionArch.=ViT-B/14, Evaluation Protocol=lin. 42026.03 | 0.458 | |
| High ResolutionArch.=ViT-B/14, Evaluation Protocol=lin. 12026.03 | 0.46 | |
| DINOv2Framework=DINOv2, Size=B, Handling Strategy=attention bias2026.02 | 0.464 | |
| layer specializationFramework=DINOv2, Size=B, Handling Strategy=attention bias2026.02 | 0.47 | |
| OpenCLIPArch.=ViT-G/14, Evaluation Protocol=lin. 42026.03 | 0.476 | |
| Low ResolutionArch.=ViT-B/14, Evaluation Protocol=lin. 12026.03 | 0.483 | |
| MAEArch.=ViT-H/14, Evaluation Protocol=lin. 42026.03 | 0.523 | |
| OpenCLIPArch.=ViT-G/14, Evaluation Protocol=lin. 12026.03 | 0.537 | |
| DINOArch.=ViT-B/8, Evaluation Protocol=lin. 42026.03 | 0.541 | |
| MAEArch.=ViT-H/14, Evaluation Protocol=lin. 12026.03 | 0.545 | |
| DINOArch.=ViT-B/8, Evaluation Protocol=lin. 12026.03 | 0.553 | |
| layer specializationFramework=DeiT-III, Size=B, Handling Strategy=attention bias2026.02 | 0.809 | |
| DeiT-IIIFramework=DeiT-III, Size=B, Handling Strategy=attention bias2026.02 | 0.823 |