Boundary Loss Ablation for Full-Resolution Cityscapes Segmentation: When Dice Helps and When It Doesn't

Companion code, configs, per-epoch metrics, figure-generation script, and paper source for a controlled 2x2 loss ablation on Cityscapes at native resolution (1024x2048) with ConvNeXt-V2-Base + UPerNet. Four loss configurations (A: CE; B: CE+Dice; C: CE+Dice+Boundary; D: CE+Boundary) trained for 160 epochs across three random seeds each. At epoch 10, the joint formulation C clearly dominates (+2.26 mIoU over B). At epoch 160 the boundary-only variant D takes the lead with 81.69 +/- 0.12 mIoU and 58.67 +/- 0.11 Boundary F1, while B retains the Trimap IoU advantage (49.02 +/- 0.28). Per-class breakdown shows D dominates large structured classes (wall +3.88, truck +5.29 vs B, bus +2.37) while B preserves thin signal-rich classes. The most actionable finding is methodological: short-epoch ablations are systematically misleading on this task. Code MIT-licensed. Paper PDF, per-run CSV exports, and the regeneration script are included. Companion landing page with the paper inline at https://guillaume-cassez.fr/voiture-autonome/cityscapes/boundary-loss-kervadec/.