Evaluation of Hierarchical Watersheds.

Area of Science:

• Computer Vision
• Image Analysis
• Morphological Segmentation

Background:

• Hierarchies of morphological segmentations are crucial for image analysis.
• Existing methods for hierarchy construction can be complex.
• Evaluating these hierarchies requires a comprehensive framework.

Purpose of the Study:

• To understand the practical features of quasi-flat zones hierarchy and watershed hierarchies.
• To evaluate their potential for natural image analysis.
• To propose a novel evaluation framework for image segmentation hierarchies.

Main Methods:

• Developed a novel evaluation framework assessing region/contour precision, cut extraction quality, and semantic object representation.
• Assessed and optimized hierarchies using pre- and post-processing steps.
• Utilized a state-of-the-art contour detector with watershed hierarchies.

Main Results:

• Watershed hierarchies, with a contour detector, are competitive with complex state-of-the-art hierarchy construction methods.
• A novel watershed hierarchy based on extinction value (number of parent nodes) outperformed others.
• The proposed framework identified optimal hierarchies for segmentation.

Conclusions:

• Watershed hierarchies are valuable for computer vision tasks due to their efficiency and global optimality properties.
• The novel evaluation framework effectively assesses and optimizes morphological segmentation hierarchies.
• Optimized watershed hierarchies show strong potential for natural image segmentation.