RGC-Mono: Ratio-Guided Geometry Calibration for Monocular 3D Object Detection
Monocular 3D detection estimates metric depth from ambiguous single-view geometry. RGC-Mono calibrates a perspective-derived depth anchor with a dimensionless log-ratio residual. Scale-Normalized Projection Ratio Residual (SPRR) represents multiplicative projection discrepancy, Geometry Sensitivity-Aware Uncertainty (GSAU) weights depth regression during training, and Orientation-Visibility Conditioned Calibration (OVCC) uses pose and visibility cues to condition the correction. With a shared pretrained MonoDGP checkpoint and 250 additional training epochs, three KITTI Car validation runs obtain Moderate AP3D|R40 of 22.91±0.14% for matched MonoDGP, 23.26±0.12% for ordinal-OVCC RGC-Mono, and 23.36±0.08% for categorical OVCC. The respective mean gains are 0.35 and 0.45 percentage points. A capacity-matched SPRR–additive comparison gives a mean difference of 0.10 points with a confidence interval spanning zero, locating the detection benefit in the integrated configuration. The original ordinal checkpoint reaches 19.49% Moderate AP3D on KITTI test under its reported protocol. The framework provides a compact way to combine relative correction, geometric uncertainty, and visibility context for car depth localization.
Authors
- Ao Xu (ORCID: https://orcid.org/0009-0001-1763-2069)
- Yuhang Zhao
Institutions
- Tongji University (CN)
- Guilin University of Technology (CN)
Publication Details
- Journal
- Symmetry
- Published
- 2026-11-17
- DOI
- https://doi.org/10.3390/sym18111923
- Primary Topic
- Advanced Neural Network Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00