Weakly Supervised Building Semantic Segmentation Based on Spot-Seeds and Refinement Process.

Area of Science:

• Geospatial analysis
• Computer vision
• Machine learning

Background:

• Automatic building semantic segmentation is crucial for geospatial applications.
• Convolutional neural networks (CNNs) dominate current methods but require extensive pixel-level labels.
• The need for large labeled datasets hinders CNN-based building segmentation.

Purpose of the Study:

• To propose a novel weakly supervised framework for building segmentation.
• To generate high-quality pixel-level annotations automatically.
• To optimize segmentation networks and reduce human labeling efforts.

Main Methods:

• A superpixel segmentation algorithm predicts boundary maps.
• Superpixels-CRF, guided by spot seeds, propagates information to unlabeled regions for annotation.
• Iterative retraining of the segmentation network using refined predicted maps.

Main Results:

• The framework generates high-quality pixel-level annotations.
• A more robust segmentation network is trained.
• Significant improvement in building segmentation quality is achieved.
• Human labeling efforts are substantially reduced.

Conclusions:

• The proposed weakly supervised framework effectively addresses the data annotation challenge in building segmentation.
• It enables the training of robust segmentation networks with reduced manual effort.
• This approach offers a promising solution for large-scale geospatial applications requiring accurate building extraction.