One Scene, Two Depths: Probing Geometric Ambiguity in Monocular Foundation Models

Xiaohao Xu · Feng Xue · Xiang Li · Haowei Li · Shusheng Yang · Tianyi Zhang · Matthew Johnson-Roberson · Xiaonan Huang

Figure 1: Rethinking geometric ambiguity for 3D spatial understanding. (a) Ambiguous layered scenes can contain multiple visible surfaces along one ray, while single-depth supervision records only one biased layer. (b) This collapse turns m — FIGURE FROM THE PAPER

TL;DR

Depth foundation models pick different layers in transparent scenes, and a training-free Laplacian input transform can shift which layer some models report.

BRIEF

A single camera ray in a transparent scene can hit both a glass surface and the object behind it, yet monocular depth models output just one depth per pixel. This work builds MD-3k, a benchmark with two-layer ordinal labels, to measure which layer a model prefers. Leading depth foundation models disagree in their preferences. A training-free input transform, Laplacian Visual Prompting, can shift the layer some frozen models report, suggesting they hold alternative geometric hypotheses that standard RGB input does not reveal.

Abstract

A faithful 3D world representation should account for layered geometry, where a single camera ray may contain multiple visible and geometrically valid surfaces. Monocular depth estimation, however, reduces this structure to one scalar depth per pixel. Transparent scenes make this ambiguity measurable: the same ray can pass through foreground glass and observe the background, turning the supervised target into a convention of annotation, data, and training rather than a scene-intrinsic truth. A learned predictor exposes this convention as its depth-layer preference. We introduce MultiDepth-3k (MD-3k), a sparse two-layer ordinal benchmark for measuring depth-layer preference and multi-layer spatial relationship accuracy (ML-SRA). On MD-3k, leading depth foundation models exhibit diverse layer preferences under standard RGB input, showing that the same layered geometry can be resolved differently across models. We further find that Laplacian Visual Prompting (LVP), a training-free spectral input transformation, can substantially change the reported layer for certain frozen models. The strongest RGB/LVP pair, DAv2-L, reaches 75.5% ML-SRA. These results suggest that depth foundation models may express complementary geometric hypotheses that standard RGB inference leaves unexpressed. We invite the community to rethink depth supervision and evaluation through an ambiguity-aware lens, where multiple valid 3D interpretations are treated as geometric structure to be measured, preserved, and expressed.