Image Segmentation Using Hierarchical Merge Tree.

Area of Science:

• Computer Vision
• Image Analysis

Background:

• Image segmentation is a fundamental problem in computer vision.
• Existing methods often rely on semantic priors, limiting their general applicability.

Purpose of the Study:

• To propose a supervised hierarchical approach for object-independent image segmentation.
• To develop an efficient method for inferring globally optimal segmentations.

Main Methods:

• Utilizes oversegmentation into superpixels and a tree structure for region merging hierarchy.
• Formulates the tree structure as a constrained conditional model with an ensemble boundary classifier.
• Employs an iterative training and testing algorithm for segmentation accumulation.

Main Results:

• Achieves state-of-the-art region accuracy on six public datasets.
• Demonstrates competitive performance in semantic-prior-free image segmentation.
• The hierarchical approach effectively reduces segmentation to label assignment on tree nodes.

Conclusions:

• The proposed supervised hierarchical method offers a robust solution for general image segmentation.
• Efficient inference and segmentation accumulation lead to high accuracy.
• The approach is effective even without relying on semantic information.