Controllable stroke-based sketch synthesis from a self-organized latent space.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Controllable synthesis of free-hand sketches is hindered by a lack of labeled training data for categories and styles.
• Organizing latent coding spaces with Gaussian mixture priors can control sketch synthesis by representing patterns.
• Determining the correct number of Gaussian components is crucial for accurate pattern clustering and controllable generation.

Purpose of the Study:

• To develop an unsupervised method for automatically determining the optimal number of Gaussian components for controllable sketch synthesis.
• To address the challenge of pattern clustering and controllable generation in latent coding spaces.

Main Methods:

• Introduced Rival Penalized Competitive Learning pixel to sequence (RPCL-pix2seq) to automatically determine the Gaussian number.
• Utilized a Gaussian mixture prior over latent codes to represent categorical or stylistic patterns.
• Employed self-organization of a latent coding space to preserve structural pattern similarity.

Main Results:

• RPCL-pix2seq successfully partitions sketch codes into a stable number of clusters.
• Quantitative and qualitative experiments validate the method's effectiveness.
• The approach enables synthesis reasoning over the latent space for novel sketch generation.

Conclusions:

• RPCL-pix2seq provides a robust solution for unsupervised determination of Gaussian numbers in controllable sketch synthesis.
• The method facilitates the generation of novel and reasonable sketches by effectively managing latent representations.
• This work advances the field of controllable generative models for sketch synthesis.