Robust vehicle re-identification with hierarchical visual refinement and conditional side-information calibration

Vehicle re-identification (Re-ID) aims to match the same vehicle across distributed traffic cameras. However, reliable matching remains difficult because fine-grained vehicle cues are easily degraded by occlusion, blur, illumination changes, and background clutter. Camera and viewpoint labels may also be incomplete or incompatible across datasets. To address these challenges, this study proposes SCGF-Net, a vehicle Re-ID framework with hierarchical visual refinement and conditional side-information calibration. SCGF-Net first strengthens middle-level local cues before deep semantic abstraction. It then propagates the resulting evidence to guide high-level gated fusion. Conditional residual calibration is applied when compatible camera or viewpoint information is available. This design preserves discriminative vehicle details, suppresses identity-irrelevant responses, and reduces reliance on complete auxiliary annotations. On VeRi-776, SCGF-Net achieved 85.9% mAP, 98.4% Rank-1, and 99.2% Rank-5, outperforming the Swin Transformer baseline by 7.0, 4.2, and 1.6% points, respectively. In the strict VeRi-776-to-VeRi-Wild transfer setting, SCGF-Net improved mAP, Rank-1, and Rank-5 over the Swin Transformer baseline by 2.6, 3.4, and 3.9% points, respectively. Ablation studies, controlled metadata perturbation experiments, and attention visualizations further verify the effectiveness and operational boundaries of the proposed components. These results support the effectiveness of SCGF-Net under dataset shift and degraded metadata conditions.