Improving Thing Segmentation for USV Panoptic Scene Understanding with Detector-Guided Class-Specific Refinement

HyewonChun, JO WONWOO, Sangmunlee

Keywords: Panoptic Segmentation, Maritime Computer Vision, Scene Understanding, Instance Segmentation

Abstract: Panoptic segmentation is well suited to unmanned surface vehicle (USV) scene understanding because safe navigation requires both dense parsing of background regions and explicit reasoning about obstacle instances. Maritime panoptic benchmarks remain challenging because thing classes are often small, sparse, and long-tailed, whereas stuff classes dominate image area. We present a practical hybrid system for the LaRS maritime panoptic benchmark that combines a refined MaskDINO panoptic ensemble with an RF-DETR-Seg Medium thing booster and class-specific GrabCut refinement during fusion, while remaining feasible to develop on a single workstation GPU. Our development analysis suggests that much of the remaining room for improvement on LaRS lies in thing segmentation rather than stuff parsing. Accordingly, we focus on panoptic-branch refinement, detector-guided fusion, and targeted boundary refinement for row boats, paddle boards, buoys, swimmers, and selected rare classes. On the local development split used for method design, our system improves panoptic quality from 37.50 to 45.79 and raises thing panoptic quality from 17.47 to 28.49 while maintaining stuff quality at 91.92. On the official hidden-test benchmark, the final anonymous submission achieves 42.6 PQ, 24.2 P Qth , 91.8 P Qst , 51.1 RQ, 71.0 SQ, and 54.4 F1. Taken together, these results suggest that detector-guided, class-specific refinement offers a practical way to improve maritime panoptic segmentation under single-workstation-GPU constraints.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 18