A Benchmark for Presence-Uncertain Few-Shot Segmentation in Real-World Remote Sensing Scenarios

Few-shot segmentation (FSS) is typically evaluated under a positive-class assumption, where the reference class is guaranteed to appear in every target image. This assumption is particularly fragile in remote sensing imagery, where large-scale scenes are processed in a patch-wise manner and target objects are often entirely absent. To better reflect real-world application scenarios, evaluation must consider not only where to segment, but also whether segmentation should be performed at all. To address this gap, we introduce the Presence-Uncertain Few-Shot Segmentation (PU-FSS) Benchmark, centered on the aerial imagery dataset iSAID, which naturally exhibits presence uncertainty due to large-scale scenes. We further include standard natural image datasets to verify that the proposed evaluation protocol is not limited to a single domain. We also propose a lightweight training- free prototype-debiasing (PD) module that suppresses background bias in similarity computation, enabling reliable abstention when the reference class is absent. The proposed PU-FSS provides an evaluation protocol aligned with the requirements of large-scale remote sensing image analysis. Our results show that the high performance of training-free FSS models under the positive-class assumption does not translate to reliability in real-world settings, and that prototype debiasing provides a practical means to evaluate such models under presence uncertainty, particularly in remote sensing imagery. Our code and dataset are available at PUFSS.