Related Experiment Videos
Multi-Dimensional Quality Assessment for Single-Image-to-3D Contents: Dataset and Model
Summary
Researchers developed the first subjective database (AIGC-SI23DCQA) for evaluating AI-generated 3D content from single images. They also proposed I3DQA, a novel objective quality assessment method, outperforming existing approaches.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Multimedia Processing
Background:
- AI-generated multimedia content, including 3D models, is rapidly advancing.
- Quality evaluation for 2D AI content is established, but assessing single-image-to-3D content quality is underexplored.
Purpose of the Study:
- To establish the first comprehensive subjective evaluation database for single-image-to-3D content quality.
- To benchmark existing quality assessment methods for this task.
- To propose a novel objective quality assessment method for single-image-to-3D content.
Main Methods:
- Created AIGC-SI23DCQA database with 100 realistic, 100 AI-generated, and 100 CG input images.
- Generated 1,500 3D contents using five algorithms and collected 94,500 annotations on texture fidelity, shape accuracy, and overall quality.
- Developed I3DQA, an objective method using source image features, projected video, patches, and LMM features integrated via symmetric transformer blocks.
Main Results:
- Existing quality assessment methods show limitations for single-image-to-3D content.
- The proposed I3DQA method demonstrates superior performance in objective quality assessment.
- Experiments validate the effectiveness of I3DQA's components and its overall efficacy.
Conclusions:
- The AIGC-SI23DCQA database provides a foundational resource for 3D content quality assessment research.
- The I3DQA method offers a robust framework for effective single-image-to-3D content quality evaluation.
- This work advances the field of AI-generated 3D content quality assessment.
