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

Area of Science:

• Computer Vision
• Remote Sensing
• Geospatial Analysis

Background:

• Building instance segmentation in remote sensing is crucial for urban management, disaster assessment, and land-cover monitoring.
• Challenges include variations in building scale, dense distribution, complex backgrounds, shadows, and occlusions, impacting accuracy, boundary recovery, and efficiency.
• Existing benchmarks lack a unified, multidimensional evaluation covering diverse aspects like accuracy, speed, and robustness.

Purpose of the Study:

• To establish a unified, multidimensional benchmark for evaluating remote sensing building instance segmentation models.
• To compare the performance of leading mask-predicting models (YOLOv8-seg, YOLOv11-seg, YOLO26-seg, Mask R-CNN) and auxiliary detection baselines (RT-DETR).
• To assess model transferability across datasets and robustness against degradation factors like shadows, occlusions, and blur.

Main Methods:

• Implemented a consistent training and evaluation framework for mask-predicting instance segmentation models.
• Included benchmark metrics for bounding-box detection, mask-based segmentation, inference efficiency, model complexity, and training behavior.
• Conducted zero-shot cross-dataset testing (WHU-to-Inria), in-domain training/testing with varied initializations, and controlled degradation tests.

Main Results:

• High-capacity YOLO-seg models demonstrated strong competitiveness, with YOLOv11x-seg achieving top mask-based accuracy and YOLOv11m-seg offering a superior balance of accuracy, speed, and complexity.
• Zero-shot WHU-to-Inria testing highlighted a significant domain shift, but in-domain training enabled YOLO-seg models to regain competitive performance.
• YOLOv11x-seg exhibited greater robustness to shadow/occlusion compared to Gaussian blur, indicating differential sensitivity to image degradations.

Conclusions:

• The developed benchmark provides crucial evidence for selecting remote sensing building instance segmentation models based on specific deployment needs (accuracy vs. efficiency).
• YOLO-seg models, particularly YOLOv11x-seg and YOLOv11m-seg, are highly competitive and adaptable for various remote sensing applications.
• Understanding domain shift and degradation robustness is essential for reliable real-world deployment of building instance segmentation systems.