Toward Deep Universal Sketch Perceptual Grouper.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Machine Learning

Background:

• Human free-hand sketches naturally exhibit Gestalt grouping principles.
• Existing methods lack a universal approach for grouping sketch components across diverse categories and styles.
• A significant barrier is the absence of large-scale, annotated sketch datasets for perceptual grouping.

Purpose of the Study:

• To develop the first universal sketch perceptual grouper capable of segmenting strokes into semantically meaningful parts.
• To create the largest dataset for sketch perceptual grouping to facilitate research.
• To advance sketch analysis tasks through improved grouping capabilities.

Main Methods:

• Contribution of the largest sketch perceptual grouping dataset (20,000 sketches, 25 categories).
• Proposal of a novel deep perceptual grouping model utilizing both generative and discriminative losses.
• Training the model to ensure generalization (generative loss) and consistency (discriminative loss).

Main Results:

• The proposed grouper significantly outperforms existing state-of-the-art methods.
• The model demonstrates strong generalization across various sketch categories and drawing styles.
• The grouper achieves high local and global grouping consistency.

Conclusions:

• The developed universal sketch perceptual grouper is effective and generalizes well.
• The novel deep learning model and dataset advance the field of sketch analysis.
• The grouper shows utility in downstream tasks like semantic segmentation and retrieval.