Answers are grounded in this paper and cite their sources.
In extremely sparse remote sensing scenes, buildings often appear as isolated small-scale targets embedded in large areas of farmland, bare land, and natural vegetation. Their extremely low area ratio and severe background interference make existing dense prediction-based building extraction methods prone to missed detections and false alarms. To address this issue, we propose ActiveScout Agent, an active localization-and-focusing agent for isolated small-scale building extraction in extremely sparse scenes. Rather than designing another segmentation backbone, ActiveScout reformulates isolated small-scale building extraction from a passive dense prediction problem into a sequential decision-making problem of “active localization and focusing followed by local fine-grained extraction” from the perspectives of task reformulation and active perception. It can also serve as a general front-end active perception module that can be coupled with different types of segmentation models. Specifically, the agent performs a coarse-to-fine search in continuous image space through translation, scaling, and termination actions, progressively narrowing the search region from global scanning to local focusing. To improve training under extremely sparse conditions, we further design a continuous reward based on geometric distance and target coverage, a scale-triggered dual-path local feature update strategy, and an optimization scheme that combines curriculum learning with collaborative training. Experiments on three constructed extremely sparse single-target benchmarks derived from the WHU Building Dataset, the Inria Aerial Image Labeling Dataset, and the Ganyu Building Dataset show that ActiveScout improves the average Intersection over Union of U-Net, DeepLabV3+, SegFormer, Atlas, Samba, and UNetMamba by 8.75%, 11.27%, 4.28%, 8.62%, 3.65%, and 3.99%, respectively.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: