Answers are grounded in this paper and cite their sources.
Recent monocular depth estimators show strong zero-shot generalization, but many remain unreliable under strict metric evaluation: relative-depth foundation models often need test-time scale/shift alignment, while metric predictors can suffer scale drift and blurred depth discontinuities under domain shift. We present FuseDepth, a prior-guided framework for zero-shot metric monocular depth estimation that injects boundary and geometric cues into a frozen foundation-depth backbone. FuseDepth adapts Depth Anything v2 with lightweight LoRA modules and a boundary-regularized local metric-scale field to obtain a coarse metric prediction. It then refines this prediction using a matting-derived surface-interaction graph whose nodes represent detected surface hypotheses and whose edges encode overlap, boundary adjacency, and contour proximity to define a soft occlusion jump set. Refinement is controlled by a learned prior-reliability gate that arbitrates between metric scale, contour evidence, and monocular surface-normal priors, rather than simply concatenating external features. We evaluate FuseDepth on unseen SYNS, DDAD, and DIODE benchmarks against relative and metric baselines under a strict target-domain zero-shot protocol with no target calibration. Across datasets, FuseDepth improves AbsRel and SILog and reduces per-image log-scale bias. Boundary-specific evaluation now covers SYNS, DDAD, and both DIODE splits, where FuseDepth consistently improves RGB-edge alignment and depth-edge F1 while DepthPro retains a small DBE or precision advantage on some splits. A target-affine oracle sensitivity analysis shows that DAv2 and Marigold improve when target ground truth is allowed for calibration, yet FuseDepth without target calibration retains lower AbsRel and SILog and higher δ 1 on every evaluated target. On DDAD, the gain is reproduced across three independently trained seeds and across all three geographic subsets. Mechanism-isolation, natural-prior-quality, wrong-prior, graph-sensitivity, and deployment-resource analyses further characterize when the structured prior fusion helps and what computational cost it incurs.
No takes yet. Share an insight, caveat, or question.
Hossein Javidnia (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: