Split-GCN: Effective Interactive Annotation for Segmentation of Disconnected Instance

Area of Science:

• Computer Vision
• Machine Learning
• Image Segmentation

Background:

• Manual annotation of object boundaries is costly and time-consuming.
• Polygon-based annotation methods with human interaction show promise but struggle with disconnected object components due to fixed topology.
• Existing methods face challenges in precisely predicting object boundaries, especially for complex shapes.

Purpose of the Study:

• To introduce Split-GCN, a novel architecture for accurate object boundary annotation.
• To address the limitation of predicting disconnected components in polygon-based annotation.
• To enhance the precision of vertex movement towards object boundaries using directional information.

Main Methods:

• Developed Split-GCN, a novel architecture combining polygon-based methods with a self-attention mechanism.
• Incorporated directional information to guide precise vertex movement along object boundaries.
• Transformed initial polygon topology using context exchange to model vertex dependencies and predict disconnected components.

Main Results:

• Split-GCN successfully predicts disconnected components within objects.
• The model demonstrates competitive performance against state-of-the-art models on the Cityscapes dataset.
• Achieved superior performance compared to baseline models on Cityscapes and showed strong generalization across four cross-domain datasets.

Conclusions:

• Split-GCN offers a significant advancement in polygon-based object boundary annotation.
• The architecture effectively handles disconnected components, a key limitation of prior methods.
• The model's strong performance and generalization ability highlight its potential for real-world applications in image segmentation.