VGGT-Metric: A Metric Depth Estimation Model Based on VGGT

• 1、School of Computer Science(National Pilot Software Engineering School), Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：Pure vision depth estimation plays a vital role in fields such as autonomous driving and 3D perception. The recently proposed Visual Geometry Grounded Transformer (VGGT) performs excellently in scene reconstruction and relative depth estimation tasks. However, due to its lack of absolute physical scale awareness, it struggles to directly output the metric depth required for autonomous driving planning and control. To address this scale ambiguity challenge, this paper proposes an end-to-end metric depth estimation model named VGGT-Metric. Building upon the original VGGT architecture, this model innovatively introduces a learnable scale token. Specifically, by fully leveraging the known camera calibration priors in autonomous driving scenarios, the camera intrinsic and extrinsic parameters are mapped into feature embeddings via a Multi-Layer Perceptron (MLP) and injected into the scale token. Subsequently, this token, along with image features extracted by DINOv2, is fed into alternating frame and global Transformer encoders, achieving thorough interaction between global scale information and local visual features through self-attention mechanisms. In the decoding phase, an independent scale prediction branch is designed to specifically decode the scale token and output a global scale factor, which is then multiplied by the relative depth map to achieve entirely post-processing-free metric depth recovery. Experimental results on the large-scale nuScenes autonomous driving dataset demonstrate that, compared to baseline models, our method achieves significant performance improvements on key metrics such as absolute relative error (Abs Rel), effectively mitigating the scale drift problem in pure vision depth estimation.

Keywords： metric depth estimation Visual Geometry Grounded Transformer scale recovery scale modeling autonomous driving