Advancing Zero-Shot Digital Human Quality Assessment Through Text-Prompted Evaluation.

Area of Science:

• Computer Vision
• Digital Media Quality Assessment
• Human-Computer Interaction

Background:

• Digital humans are increasingly used across various fields, highlighting the need for quality assessment.
• Existing digital human quality assessment (DHQA) databases are insufficient for comprehensive research.

Purpose of the Study:

• To introduce the SJTU-H3D database, a subjective quality assessment resource for full-body digital humans.
• To propose a novel zero-shot DHQA method for no-reference (NR) scenarios.

Main Methods:

• Created SJTU-H3D with 40 reference and 1,120 distorted digital humans across seven distortion types.
• Developed a zero-shot DHQA approach integrating semantic (CLIP), distortion (NIQE), and geometry (dihedral angles) features.
• Introduced the Digital Human Quality Index (DHQI) by aggregating these features.

Main Results:

• The SJTU-H3D database provides a benchmark for evaluating and refining DHQA algorithms.
• The proposed zero-shot DHQA method demonstrates significant performance improvements.
• The DHQI serves as a robust baseline for various DHQA tasks.

Conclusions:

• The SJTU-H3D database and the proposed DHQI method advance the field of digital human quality assessment.
• The zero-shot approach ensures generalization and mitigates database bias in DHQA.
• The developed resources facilitate further research and development in digital human technologies.