Fast Sketch Segmentation and Labeling With Deep Learning.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Modeling

Background:

• Automatic decomposition of sketched objects into meaningful parts is challenging.
• Existing methods often require extensive annotated sketch data for training.

Purpose of the Study:

• To develop an efficient deep learning method for automatic semantic decomposition of freehand sketches.
• To enable accurate part segmentation and labeling of sketches without large annotated datasets.

Main Methods:

• A deep neural network trained to transfer 3D model segmentations and labelings to 2D sketches.
• Input: binary image of a sketched object. Output: segmentation map with per-pixel labelings.
• Postprocessing using multilabel graph cuts for refinement.

Main Results:

• The method achieves high segmentation and labeling accuracy on sketch datasets.
• Demonstrates significant speed improvement over state-of-the-art methods.
• Successfully integrated into a sketch-based 3D modeling application.

Conclusions:

• The proposed deep learning approach offers an efficient and accurate solution for sketch decomposition.
• Enables practical applications like automated 3D model generation from sketches.
• Reduces the need for large annotated sketch datasets in training.