RefineAny3D Uses VLM Visual Alignment to Refine Monocular 3D Detection Without Numbers

kwangmoo_yi · x · 2026-08-12

RefineAny3D introduces a novel depth refinement method for monocular 3D object detection. The authors found that current depth foundation models, despite strong zero-shot generalization, lack the object-level precision required for accurate 3D detection.

The core insight is that depth error has a direct visual signature in image space: a correctly placed 3D bounding box projects tightly around the object, while a too-far box projects too small and a too-close box projects too large.

By framing depth refinement as a visual alignment problem rather than metric regression, the authors fine-tune a Vision-Language Model (VLM). They extend the VLM's vocabulary with action tokens to replace numerical depth outputs with categorical decisions, supervised on a large-scale chain-of-thought dataset.

Related event: RefineAny3D Boosts Monocular 3D Detection via Fine-tuned VLMs(2 posts)→

Original post →

More from Research

Research channel →