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

Companion code, configs, per-class metrics for twelve runs, figure/statistics scripts, 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), each trained for 160 epochs across three random seeds. At epoch 10 the joint variant C leads (+2.26 mIoU over B in the mean); at epoch 160 the boundary-only variant D has the highest mean mIoU (81.69 +/- 0.25) and Boundary F1 (77.32 +/- 0.13). mIoU significance uses a paired image-bootstrap over the 500 val images as the pre-specified primary test (3-seed paired-t as robustness): under Holm, D beats the field-standard CE+Dice baseline B (p=0.032) and plain CE A (p=0.043), while D > C is significant only at the seed level (p=0.042). The two contour metrics split along the Dice axis (non-Dice A, D lead Boundary F1; Dice B, C lead Trimap IoU by ~1.6). The most actionable finding is methodological -- short-epoch ablations reverse the converged ranking. A connected-component consensus filter adapted from BRATS is fully characterised: the canonical pairing C-veto-B (the boundary-with-Dice variant C vetoed by the CE+Dice baseline B -- the direct BRATS analogue) prunes -18.7% of fragments at no mIoU cost and no meaningful boundary-quality cost (a pure spatial-coherence gain). v0.2.0 corrected two estimator issues from v0.1.0 (per-class Boundary F1 / Trimap IoU instead of road-vs-rest; Student-t CIs with Holm-corrected significance); v0.2.1 quantified the consensus filter; v0.2.2 finalises it with the paired image-bootstrap as the primary mIoU test and the canonical C-veto-B characterisation. The mIoU is validated bit-for-bit against the official cityscapesscripts. See CHANGELOG.md. Code MIT-licensed. Companion landing page with the paper inline at https://guillaume-cassez.fr/voiture-autonome/cityscapes/boundary-loss-kervadec/.