Learning to Aggregate Multi-Scale Context for Instance Segmentation in Remote Sensing Images.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Instance segmentation in remote sensing is crucial for civil applications.
• Existing natural image methods fail on remote sensing data due to scale variation, low contrast, and clustered objects.
• Lack of discriminative features hinders performance.

Purpose of the Study:

• To propose a novel context aggregation network (CATNet) for improved instance segmentation in remote sensing images.
• To address the challenges of scale variation, low contrast, and clustered object distributions.
• To enhance feature extraction for more accurate per-pixel labeling.

Main Methods:

• Developed CATNet, a context aggregation network.
• Integrated three lightweight modules: Dense Feature Pyramid Network (DenseFPN), Spatial Context Pyramid (SCP), and Hierarchical Region of Interest Extractor (HRoIE).
• DenseFPN enables flexible information flow; SCP uses attention for global spatial context; HRoIE generates adaptive RoI features.

Main Results:

• CATNet significantly outperforms state-of-the-art methods on iSAID, DIOR, NWPU VHR-10, and HRSID datasets.
• The proposed approach achieves superior performance with comparable computational costs.
• Demonstrated effectiveness in aggregating global visual context across feature, spatial, and instance domains.

Conclusions:

• CATNet effectively addresses key challenges in remote sensing instance segmentation.
• The proposed modules enhance feature representation and context aggregation.
• CATNet offers a robust and efficient solution for per-pixel object labeling in remote sensing imagery.