DepthEvidence is a 4‑billion‑parameter multimodal language model that unifies metric depth prediction with geometric reasoning within a single language generation pipeline. The decoder is conditioned on camera parameters, allowing it to predict full‑resolution metric depth maps from multi‑scale visual features and high‑resolution RGB refinement. A dense‑to‑language interface then converts the predicted depths and decoder features into continuous geometry tokens anchored to object identifiers, which serve as evidence for subsequent language generation.
Geometric supervision is designed so that metric information remains recoverable both before and after language‑context interaction, preserving numerical content throughout reasoning. Instruction tuning equips the model with abilities such as object measurement, relative comparison, and compositional reasoning. To benchmark spatial‑numeric reasoning, the authors introduce the Depth‑VQA suite, covering object‑depth queries, relative comparisons, and decisions that combine spatial and numeric constraints.
Across nine public datasets, DepthEvidence achieves the highest average dense $\delta_1$ score among evaluated methods and matches the performance of specialized depth estimators. It leads on instance‑level metric depth estimation and on both the relative and metric reasoning tracks, while largely retaining the original VQA performance and improving spatial understanding compared to the base model.
Review: DepthEvidence demonstrates the feasibility of embedding precise visual geometry directly into language models, maintaining numerical fidelity while enabling flexible textual reasoning. This work opens a promising direction for multimodal AI systems that require robust spatial cognition.