A Unified Multidimensional Benchmark and Multi-Dataset Evaluation of YOLO-Based Models for Remote Sensing Building Instance Segmentation

Building instance segmentation in remote sensing imagery supports applications such as urban management, disaster assessment, 3D urban modeling, and land-cover monitoring. However, variations in building scale, dense spatial distribution, complex background textures, shadows, and occlusions make it difficult to balance segmentation accuracy, boundary recovery, inference efficiency, and deployment cost. This study establishes a unified multidimensional benchmark for remote sensing building instance segmentation. The primary benchmark evaluates mask-predicting instance segmentation models, including YOLOv8-seg, YOLOv11-seg, YOLO26-seg, and Mask R-CNN, under consistent training and evaluation settings. RT-DETR-l and RT-DETR-x are retained only as auxiliary detection-only Transformer baselines because they do not output instance masks in the implemented setting. The benchmark covers bounding-box detection, mask-based segmentation, inference efficiency, model complexity, training behavior, and qualitative visualization. To assess cross-dataset transferability and degradation-specific robustness beyond a single dataset, we further conduct zero-shot WHU-to-Inria testing, independent Inria training/testing with different initialization strategies, and controlled degradation tests involving shadow/occlusion and Gaussian blur. Results on WHU and Inria show that high-capacity YOLO-seg models are competitive among the evaluated mask-predicting models. Under the current experimental settings, YOLOv11x-seg achieves the highest or near-highest mask-based accuracy, whereas YOLOv11m-seg provides a favorable balance between accuracy, speed, and complexity. The zero-shot WHU-to-Inria test reveals a clear domain shift, while the Inria in-domain experiments indicate that high-capacity YOLO-seg models recover competitive performance after target-domain training. The controlled degradation tests show a smaller performance drop under shadow/occlusion than under Gaussian blur for YOLOv11x-seg. These findings provide benchmark-specific evidence for selecting remote sensing building instance segmentation models under accuracy-oriented and efficiency-oriented deployment requirements.