Research map / 3D vision / Depth and stereo
Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer
Rene Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler B, Vladlen Koltun
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Pareto-optimal multi-objective mixing instead of naive mixing (all five datasets) (ablation) | MIX 5 · Mean over 6 zero-shot test sets | mean relative performance [%]: 19.48 -> 22.40 | Improved |
| Add MV to MIX 1 with naive mixing (MIX 1 to MIX 2) (ablation) | naive mixing · Mean over 6 zero-shot test sets | mean relative performance [%]: 18.26 -> 17.34 | Hurt |
| Add MV to MIX 1 with Pareto mixing (MIX 1 to MIX 2) (ablation) | Pareto mixing · Mean over 6 zero-shot test sets | mean relative performance [%]: 16.56 -> 19.75 | Improved |
| Add MD to MIX 2 with Pareto mixing (MIX 3) (ablation) | Pareto mixing · Mean over 6 zero-shot test sets | mean relative performance [%]: 19.75 -> 20.79 | Improved |
| Add WS to MIX 3 (MIX 5) with Pareto mixing (ablation) | Pareto mixing · Mean over 6 zero-shot test sets | mean relative performance [%]: 20.79 -> 22.40 | Improved |
| Add 3D Movies (MV) to MIX 4 (giving MIX 5), Pareto mixing (ablation) | Pareto mixing · Mean over 6 zero-shot test sets | mean relative performance [%]: 21.6 -> 22.40 | Improved |
| Fine-tune on single dataset DL instead of RW only (ablation) | RW baseline · Mean over 6 zero-shot test sets | mean relative performance [%]: -10.20 | Hurt |
| Fine-tune on single dataset WS instead of RW only (ablation) | RW baseline · Mean over 6 zero-shot test sets | mean relative performance [%]: -21.37 | Hurt |
| Swap ResNet-50 encoder for ResNeXt-101 pretrained on weakly-supervised data (ablation) | MiDaS depth network · Validation sets (RW, MD, MV) | relative performance: up to 15 % relative improvement | Improved |
| Initialize encoder with ImageNet pretraining vs random init (ablation) | ResNet-50 encoder · Validation sets | relative performance: 35\% worse | Improved |
| Use trimmed MAE scale-shift-invariant loss L_ssitrim + L_reg (ablation) | depth network trained on RW · Validation sets | relative performance | Improved |
| Use full (ResNeXt-101-WSL) model instead of small ResNet-50 model | Ours · Mean over 6 zero-shot test sets | mean relative performance [%]: -16.0 | Improved |
Closest papers by meaning
- Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data (CVPR 2024 · similarity 0.90)
- BetterDepth: Plug-and-Play Diffusion Refiner for Zero-Shot Monocular Depth Estimation (NeurIPS 2024 · similarity 0.89)
- DepthFM: Fast Generative Monocular Depth Estimation with Flow Matching (AAAI 2025 · similarity 0.89)
- MonoInstance: Enhancing Monocular Priors via Multi-view Instance Alignment for Neural Rendering and Reconstruction (CVPR 2025 · similarity 0.89)
- DME: Unveiling the Bias for Better Generalized Monocular Depth Estimation (AAAI 2024 · similarity 0.89)
- Depth Pro: Sharp Monocular Metric Depth in Less Than a Second (ICLR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
