Towards Ambiguity-Free Spatial Foundation Model: Rethinking and Decoupling Depth Ambiguity
Xiaohao Xu · Feng Xue · Xiang Li · Haowei Li · Shusheng Yang · Tianyi Zhang · Matthew Johnson-Roberson · Xiaonan Huang
![Figure 2: Unlocking hidden depth with Laplacian Visual Prompting across diverse baselines [depth_anything, yang2024depth, marigold, dpt]. Each case includes the RGB input, estimated depth from RGB, Laplacian input, estimated hidden depth fr](/_next/image?url=%2Fpapers%2Ftowards-ambiguity-free-spatial-foundation-model%2Fimage.webp&w=3840&q=75&dpl=dpl_8ygxTXqcm6LhJeKCMTS1uNCnV7Kk)
TL;DR
A training-free Laplacian image prompt lets pretrained depth models reveal a second depth layer behind glass, evaluated on the new MD-3k benchmark.
BRIEF
Depth estimators usually return one depth value per pixel, which fails when a pixel shows both a transparent surface and the objects behind it. The authors introduce MD-3k, a benchmark that labels near/far relationships across multiple layers in ambiguous regions. They also propose Laplacian Visual Prompting, which feeds a Laplacian-transformed image to a pretrained depth model to recover a second, hidden depth layer without retraining. Combined with the standard RGB estimate, this gives multi-layer depth. The authors report benefits for depth-conditioned image generation, 3D spatial reasoning, and video depth.
Abstract
Depth ambiguity is a fundamental challenge in spatial scene understanding, especially in transparent scenes where single-depth estimates fail to capture full 3D structure. Existing models, limited to deterministic predictions, overlook real-world multi-layer depth. To address this, we introduce a paradigm shift from single-prediction to multi-hypothesis spatial foundation models. We first present MD-3k, a benchmark exposing depth biases in expert and foundational models through multi-layer spatial relationship labels and new metrics. To resolve depth ambiguity, we propose Laplacian Visual Prompting (LVP), a training-free spectral prompting technique that extracts hidden depth from pre-trained models via Laplacian-transformed RGB inputs. By integrating LVP-inferred depth with standard RGB-based estimates, our approach elicits multi-layer depth without model retraining. Extensive experiments validate the effectiveness of LVP in zero-shot multi-layer depth estimation, unlocking more robust and comprehensive geometry-conditioned visual generation, 3D-grounded spatial reasoning, and temporally consistent video-level depth inference. Our benchmark and code will be available at https://github.com/Xiaohao-Xu/Ambiguity-in-Space.