One Sketch for All: One-Shot Personalized Sketch Segmentation

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Sketch segmentation is crucial for fine-grained sketch analysis.
• Existing methods often require large labeled datasets, limiting personalization.
• Personalized segmentation enables adapting models to individual user styles and part annotations.

Purpose of the Study:

• To develop the first one-shot personalized sketch segmentation method.
• To enable segmentation of sketches from a single exemplar sketch with part annotations.
• To ensure preservation of part semantics and robustness to style variations.

Main Methods:

• A sketch-specific hierarchical deformation network is proposed.
• A graph convolutional network encodes multi-level sketch strokes.
• Hierarchical deformation estimates rigid-body transformations and stroke-wise adjustments guided by unsupervised keypoint learning.

Main Results:

• The proposed method outperforms state-of-the-art segmentation and perceptual grouping baselines by over 10%.
• It also surpasses few-shot 3D shape segmentation methods in the one-shot setting.
• Ablation studies confirm robustness to changes in part semantics and style differences.

Conclusions:

• The developed one-shot personalized sketch segmentation method effectively preserves semantic information.
• The approach demonstrates significant improvements and robustness, enabling personalized fine-grained sketch analysis.
• This work paves the way for more adaptable and user-centric sketch understanding systems.