Research map / Label-efficient and robust learning / Recognition and long-tail
DME: Unveiling the Bias for Better Generalized Monocular Depth Estimation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Distance-aware Multi-Expert (DME), a monocular depth model whose experts each cover a depth range chosen by frequency, combined by a pixel-level router, to improve generalization.
Problem
Monocular depth estimation generalizes poorly, which the paper attributes to long-tailed, imbalanced depth distributions in training data (across datasets and within individual images) and a Simulation Correlation effect.
Approach
Divide the depth range into intervals (1-3, 3-10, 10-25, 25-80 m) with one expert per interval so each expert sees a more uniform depth distribution, and use a pixel-level routing module to softly combine expert outputs, instead of image-level scene routing as in ZoeDepth.
Key results
- DME reaches d1 0.956 and RMSE 0.268 on NYUD v2 (vs ZoeDepth 0.953 / 0.277) and d1 0.980 with RMSE 1.905 on KITTI (vs ZoeDepth 0.967 / 2.290).
- On zero-shot DIODE Indoor, DME achieves d1 0.479 vs ZoeDepth 0.386; on iBims-1 it reduces RMSE by 0.148 relative to ZoeDepth.
- Using 8 experts (DME-GT-8) lowers NYUD v2 RMSE to 0.225, and distance-aware grouping beats equal-interval grouping (RMSE 0.244 vs 0.270).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace ZoeDepth's indiscriminate distance handling with distance-aware multi-expert grouping (ablation) | ZoeDepth · NYUD v2 test | RMSE: 0.277 -> 0.244 | Helped |
| replace ZoeDepth's indiscriminate distance handling with distance-aware multi-expert grouping (ablation) | ZoeDepth · KITTI test | RMSE: 2.103 -> 1.907 | Helped |
| use equal 0-20, 20-40, 40-60, and 60-80 meter expert ranges (ablation) | ZoeDepth · NYUD v2 test | RMSE: 0.277 -> 0.270 | Helped |
| use equal 0-20, 20-40, 40-60, and 60-80 meter expert ranges (ablation) | ZoeDepth · KITTI test | RMSE: 2.103 -> 2.049 | Helped |
| reweight loss toward infrequently occurring depth values (ablation) | ZoeDepth · NYUD v2 test | RMSE: 0.277 -> 0.334 | Hurt |
| assign larger loss weights to farther depth values (ablation) | ZoeDepth · NYUD v2 test | RMSE: 0.277 -> 0.337 | Hurt |
| downsample supervision for dominant and common depth ranges (ablation) | ZoeDepth · NYUD v2 test | RMSE: 0.277 -> 0.269 | Helped |
| increase the expert count to 8 (ablation) | DME · NYUD v2 test | RMSE: 0.244 -> 0.225 | Helped |
| use DME rather than ZoeDepth | depth estimation model · NYUD v2 test | RMSE: 0.277 -> 0.268 | Helped |
| use DME rather than ZoeDepth | depth estimation model · KITTI test | RMSE: 2.290 -> 1.905 | Helped |
| use DME rather than ZoeDepth | depth estimation model · Virtual KITTI 2 | RMSE: 5.095 -> 4.433 | Helped |
| use DME rather than ZoeDepth | depth estimation model · DIML Outdoor | RMSE: 3.610 -> 3.793 | Hurt |
Surprising findings
- On the zero-shot DIML Outdoor and DIODE Outdoor sets, DME's d1 (0.199, 0.251) is below ZoeDepth's (0.292, 0.208) on DIML but not uniformly better, and its DIODE Indoor REL (0.744) is much worse than ZoeDepth's (0.331) despite higher d1.
Limitations
- Routing mechanism accuracy is insufficient and limits overall performance, per the authors.
- Depth ranges were chosen empirically from observed distributions rather than learned.
- Some zero-shot metrics are mixed relative to ZoeDepth (e.g., DIODE Indoor REL, DIML Outdoor d1).
Experimental setup
Datasets: NYUD v2 KITTI DIODE iBims-1 DIML Outdoor Virtual KITTI 2
Benchmarks: NYUD v2 KITTI DIODE Indoor DIODE Outdoor iBims-1 DIML Outdoor Virtual KITTI 2
Metrics: d1 d2 d3 RMSE log10 Abs.Rel RMSE log
Baselines: BTS AdaBins LocalBins NeWCRFs PixelFormer ZoeDepth MIM GLP Resampling Reweighting
Builds on: ZoeDepth MiDaS RIDE TADE Attention-Driven Loss
Closest papers by meaning
- Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · similarity 0.90)
- Depth Pro: Sharp Monocular Metric Depth in Less Than a Second (ICLR 2025 · similarity 0.89)
- Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data (CVPR 2024 · similarity 0.88)
- A Simple yet Universal Framework for Depth Completion (NeurIPS 2024 · similarity 0.88)
- DepthMaster: Unified Monocular Depth Estimation for Perspective and Panoramic Images (NeurIPS 2026 · similarity 0.88)
- Scalable Autoregressive Monocular Depth Estimation (CVPR 2025 · similarity 0.88)
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.