Semantic boundary refinement for traffic salient object segmentation using residual logit correction

Abstract Traffic salient object segmentation requires complete regions and reliable contours for pedestrians, vehicles, signs, and obstacles, yet driving scenes contain shadows, lane markings, reflections, occlusion, and small targets that make these two goals conflict. We study this problem as task-aware semantic boundary refinement rather than generic edge detection. The supervised boundary is explicitly defined as the morphological boundary of the binary traffic-salient mask; it therefore excludes non-salient image gradients but does not encode a hierarchy of safety importance. EdgeRefine augments DeepLabV3-ResNet50 with a zero-initialized residual correction module in mask-logit space. In an 18run study on TSOD10K with three seeds per setting, the selected consistency weight of 0.01 changes ODS from 0.4875 ± 0.0016 to 0.4906 ± 0.0026,Boundary IoU from 0.5847 ± 0.0020 to 0.5881 ± 0.0004, and edge MAE from 0.004621 ± 0.000030 to 0.004496 ± 0.000008. The same setting slightly decreases Dice and increases Chamfer distance, and most paired 95% confidence intervals for overlap metrics include zero; the gains are therefore modest rather than universally significant. A SegFormer-B0 portability study yields larger improvements in Dice, Boundary F1, and Boundary IoU, while zero-shot DUTS-TE evaluation exposes backbone-dependent transfer: the Seg- Former variant improves most region and overlap metrics but not Chamfer, whereas the DeepLabV3 variants generally degrade. Condition-wise HD95and directed-distance diagnostics show that local overlap improvements can coexist with distant false contour fragments. These results position EdgeRefine as a lightweight, model-portable boundary correction mechanism with measurable benefits and explicit geometric and cross-domain limitations.