Search papers, labs, and topics across Lattice.
This paper introduces MultiDepth-3k (MD-3k), a benchmark designed to assess depth-layer preferences and multi-layer spatial relationship accuracy in monocular depth estimation. By leveraging transparent scenes, the authors reveal that leading depth foundation models exhibit varying preferences for depth layers, indicating that different models interpret the same geometric ambiguity in distinct ways. Notably, the application of Laplacian Visual Prompting (LVP) significantly alters the reported depth layers for certain models, suggesting that current depth estimation methods may overlook valuable geometric insights.
Depth foundation models reveal surprising geometric ambiguities, with the same scene interpreted differently depending on the model and input transformation used.
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.
