Monocular depth models don't agree on where a glass surface ends and the room behind it begins.
Researchers introduced MultiDepth-3k, a benchmark of roughly 3,000 sparse two-layer depth annotations designed to expose how depth foundation models handle transparent surfaces. The core finding: when a camera ray passes through foreground glass and also captures a background surface, different models pick different layers as the "true" depth — and none of them is objectively wrong. The team also tested Laplacian Visual Prompting, a training-free spectral filter applied to input images, and found it can flip a model's layer preference without any retraining. The best-performing combination, a large variant of Depth Anything v2 with the filter applied, hit 75.5% accuracy on the new multi-layer spatial relationship metric.
The result matters because monocular depth estimation underpins a wide range of downstream applications — autonomous vehicles, augmented reality, robotic manipulation — where glass surfaces are common and geometric ambiguity is not a theoretical problem. If two depth models give contradictory outputs for the same scene and both are technically defensible, then "ground truth" in depth supervision is partly a product of annotation convention rather than physical reality.
The paper is essentially an argument that the field has been grading models on a curve it never acknowledged. Most depth benchmarks assume one correct depth per pixel; this work treats layered geometry as structure to be measured, not noise to be suppressed — a framing the field hasn't widely adopted yet.