Title: GABI: Geometry-Aware Boundary Integration for Spacecraft Segmentation

Abstract: Precise segmentation is a fundamental requirement for autonomous spacecraft, as it directly influences downstream operations involving 3D situational awareness. However, the extreme lighting environments found in space result in images with significant appearance variability, which often impedes the ability of segmentation models to generalize across different spacecraft and settings. To address this challenge, we introduce GABI, a lightweight, multi-task segmentation framework that incorporates an auxiliary head for distance-field prediction into a convolutional backbone. By offering dense geometric guidance near object edges, the distance field helps the network acquire spatially coherent representations of spacecraft structures without increasing model complexity, thereby making it suitable for onboard perception systems. We benchmarked GABI against a standard convolutional baseline and a more resource-intensive transformer-based architecture. Results from the SPARK benchmark indicate that incorporating distance-field supervision boosts the baseline’s Average Precision by as much as 5%, while delivering performance on par with transformer models. Furthermore, generalization tests show that GABI increases Average Precision by over 50% compared to the baseline. In cross-domain assessments, the compact GABI variant achieved IoU and F1-scores within 5% of the heavier transformer model, despite being roughly ten times smaller in size. Meanwhile, the larger GABI variant outperformed transformer architectures while remaining nearly three times lighter.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC