FarSeg++: Foreground-Aware Relation Network for Geospatial Object Segmentation in High Spatial Resolution Remote...

Area of Science:

• Earth Vision
• Remote Sensing
• Computer Vision

Background:

• Geospatial object segmentation in high spatial resolution (HSR) remote sensing imagery faces challenges like scale variation, high intra-class background variance, and foreground-background imbalance.
• Existing semantic segmentation methods primarily address scale variation, neglecting other critical issues in large-area Earth observation.

Purpose of the Study:

• To propose a novel foreground-aware relation network (FarSeg++) to tackle the challenges of background variance and foreground-background imbalance in HSR remote sensing imagery.
• To enhance the discrimination of foreground features and improve objectness prediction for more accurate geospatial object segmentation.

Main Methods:

• Introduced a foreground-scene relation module to leverage object-scene relationships for improved feature discrimination.
• Developed foreground-aware optimization strategies to focus training on critical foreground and hard background examples.
• Proposed a foreground-aware decoder to enhance objectness representation, addressing a key bottleneck in segmentation accuracy.

Main Results:

• FarSeg++ demonstrates superior performance compared to state-of-the-art generic semantic segmentation methods on HSR remote sensing data.
• The method effectively alleviates issues of background variance and foreground-background imbalance.
• Achieved a favorable balance between processing speed and segmentation accuracy.

Conclusions:

• FarSeg++ offers a significant advancement in geospatial object segmentation for HSR remote sensing imagery.
• The proposed foreground modeling techniques provide a robust solution to persistent segmentation challenges.
• The introduced dataset and method contribute to pushing the boundaries of objectness prediction in Earth vision tasks.