A Comprehensive Evaluation of Arbitrary Image Style Transfer Methods.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Arbitrary image style transfer (AST) has advanced significantly, but inconsistent evaluation methods hinder reliable comparisons.
• Existing AST research lacks robust objective metrics and consistent subjective feedback, impeding progress.
• The need for standardized and comprehensive evaluation frameworks is critical for advancing AST.

Purpose of the Study:

• To propose a novel multi-granularity assessment system for arbitrary image style transfer (AST).
• To establish standardized objective and subjective evaluation methods for AST.
• To facilitate reliable comparisons and drive innovation in the AST field.

Main Methods:

• Collected a fine-grained dataset with diverse image contexts and parsing information.
• Developed an online evaluation system for subjective studies using point-wise, pair-wise, and group-wise questionnaires.
• Combined objective and subjective evaluations to bridge the gap and ensure consistency.

Main Results:

• Experimental evaluation of CNN-based, flow-based, transformer-based, and diffusion-based AST methods using the proposed system.
• Demonstrated the system's capability to provide standardized measures, objective data, and detailed subjective feedback.
• Examined the consistency between objective and subjective evaluation results.

Conclusions:

• The proposed multi-granularity assessment system provides a reliable and robust foundation for evaluating AST methods.
• Standardized measures and comprehensive feedback empower researchers to make informed comparisons.
• This work facilitates informed decision-making and accelerates innovation in arbitrary image style transfer research.