Coordinated Training- and Inference-Time Boundary Optimization for Urban Scene Semantic Segmentation: An Empirical Analysis

Coordinated Training- and Inference-Time Boundary Optimization for Urban Scene Semantic Segmentation: An Empirical Analysis

Abstract—Boundary displacement remains common around road edges, thin objects, and distant road users in urban scene semantic segmentation. We examine how boundary supervision during training combines with boundary correction at inference, using DeepLabV3 Plus with a ResNet-101 backbone. Training uses active boundary loss and weight averaging across three random seeds. Inference combines horizontal-flip augmentation with SegFix at native image resolution. Matched experiments compare cross-entropy and boundary-aware training under four inference protocols, measuring regional accuracy, boundary quality, per-class performance, and computational cost. On the Cityscapes validation set, active boundary loss gives positive gains across the three DeepLabV3 Plus training seeds and the four matched inference protocols. Horizontal flipping and SegFix provide further gains, and the complete configuration reaches 72.39 percent mean intersection over union. Experiments with SegFormer-B0 also show positive boundary F-score gains from active boundary loss in all three seeds. Its complete configuration increases mean intersection over union from 47.17 to 49.10 percent, indicating that the benefit is not limited to DeepLabV3 Plus. In the fully matched DeepLabV3 Plus experiments, the boundary F-score gain from active boundary loss remains positive but is smaller after SegFix. This is consistent with complementary effects and some overlap in the errors corrected by the two methods. The results provide an empirical basis for choosing training and inference settings when boundary accuracy and computational cost are both important.

Keywords-Urban Scene Semantic Segmentation; Deeplabv3+; Active Boundary Loss; Test-Time Augmentation; Boundary Evaluation