Human Annotation Dataset
收藏资源简介:
Human Annotation Dataset是由清华大学深圳国际研究生院和浙江大学联合创建的一个标准化图像提示集,用于多视角扩散模型的评估与对齐人类偏好。该数据集通过从DALL·E和Objaverse收集的600个高质量文本提示生成1200个图像提示,并使用四种多视角扩散方法生成10200个多视角资产,最终通过20位专家的配对比较,形成了16000条有效的比较数据。该数据集的创建旨在解决现有3D生成方法在评估中与人类偏好不一致的问题,特别是在图像驱动的3D生成任务中,提供了一个公平和透明的评估环境。
The Human Annotation Dataset is a standardized image prompt set jointly developed by Tsinghua Shenzhen International Graduate School and Zhejiang University, designed for the evaluation of multi-view diffusion models and alignment with human preferences. It generates 1200 image prompts from 600 high-quality text prompts collected from DALL·E and Objaverse, produces 10200 multi-view assets via four multi-view diffusion methods, and ultimately yields 16000 valid comparison data points through pairwise comparative assessments conducted by 20 experts. This dataset is created to resolve the inconsistency between existing 3D generation methods and human preferences during evaluation, and provides a fair and transparent evaluation environment, especially for image-driven 3D generation tasks.

- 1MVReward: Better Aligning and Evaluating Multi-View Diffusion Models with Human Preferences清华大学深圳国际研究生院, 浙江大学 · 2024年



