Towards Automatic Skeleton Extraction With Skeleton Grafting.

Area of Science:

• Computer Vision
• Computational Geometry
• Shape Analysis

Background:

• Existing skeleton extraction methods struggle with preserving shape features and often require manual intervention.
• Manual tuning, such as boundary smoothing and skeleton pruning, is frequently needed for accurate shape representation.

Purpose of the Study:

• To introduce a novel, automated approach for generating visually accurate skeletons without manual tuning.
• To enhance the accuracy of skeleton-based matching through improved skeleton representation and novel matching functions.

Main Methods:

• Generating backbone and dense skeletons from shape input.
• Extending backbone branches using skeleton grafting from the dense skeleton.
• Introducing two potential functions within a high-order matching protocol to fuse graph similarities and geometric relations.

Main Results:

• The proposed method generates skeletons that effectively depict shapes at levels comparable to human perception.
• The novel potential functions significantly reduce incorrect matches in skeleton-based matching protocols.
• Experiments demonstrated the effectiveness of the approach across three high-order matching protocols.

Conclusions:

• The novel automated skeleton generation approach overcomes limitations of existing methods.
• The enhanced matching protocol improves the reliability and accuracy of skeleton-based comparisons.
• This work offers a more perceptually relevant and computationally efficient solution for skeletonization and matching.