Título: Refining Runway Instance Segmentation Masks via Post-Processing for Vision-Based Landing Systems
Autores: Júlia Santos Moura, João Paulo Lara, João Pedro Klock Ferreira, Cristiano Leite de Castro & Gabriel Lott
Resumo: Instance segmentation masks for airport runway-related elements—such as runways, thresholds, and aiming points—often exhibit irregular contours and fail to accurately capture the geometric properties of the target instances. To address this limitation, we propose a model-agnostic post-processing pipeline designed to refine the shapes of these instances using only the predicted masks. Our approach comprises three specialized pipelines tailored to each instance type, followed by alignment and inference stages to ensure geometric consistency and recover missing elements. Inspired by the Smoothing Post-processing Module (SPM) from BARS, our method improves mask shapes across varying conditions, including changes in lighting, rotation, and scale, without requiring access to the internals of the segmentation model. The refined masks maintain high overlap with the originals while exhibiting improved shape regularity, significantly reducing the need for manual annotation. This enables more efficient dataset creation and supports the development of more realistic and accurate vision-based landing systems.
Palavras-chave: instance segmentation; runway; masks post-processing; plane landing.
Páginas: 7
Código DOI: 10.21528/CBIC2025-1191908
Artigo em PDF: CBIC_2025_paper1191908.pdf
Arquivo BibTeX:
CBIC_2025_1191908.bib