RGB-Only Traffic Object Ranging without Known Camera Intrinsics

Abstract

Traffic video often lacks camera intrinsics, installation height, and surveyed scale. We study coarse metric ranging from a single RGB frame under this zero-user-calibration setting. The proposed pipeline combines GeoCalib, Depth Anything V2 Metric, robust bottom-contact sampling, and 3-D backprojection, withholding estimates with insufficient or heterogeneous depth support. We conduct a controlled 100-frame KITTI forward-view evaluation with identical detections and rules for our method, MoGe-2, and GeoCalib + Metric3D v2-S. On the 281-object common-valid subset, our method obtains 2.025 m range MAE, 10.67% WAPE, 88.97% δ1, and +0.099 m signed bias, compared with 3.838 m MAE and -3.821 m bias for MoGe-2. MoGe-2 remains slightly better for pair distance (2.550 m versus 2.706 m MAE). The results demonstrate improved common-valid absolute object ranging under the evaluated protocol, while validation on fixed roadside surveillance cameras remains an important direction for future work.