MATE-3D
收藏资源简介:
MATE-3D是由上海交通大学和密苏里大学堪萨斯城分校联合创建的一个多维度文本到3D质量评估基准。该数据集包含1280个生成的纹理网格,涵盖了八个不同的提示类别,每个样本都从四个评估维度(语义对齐、几何质量、纹理质量和整体质量)进行了详细标注。数据集的创建过程包括使用大规模语言模型生成160个提示,并通过八种流行的文本到3D生成方法生成样本。MATE-3D旨在解决现有基准在提示多样性和多维度评估方面的不足,为文本到3D生成方法的研究提供了全面的评估工具。
MATE-3D is a multi-dimensional text-to-3D quality evaluation benchmark jointly created by Shanghai Jiao Tong University and the University of Missouri-Kansas City. This dataset contains 1,280 generated textured meshes, covering eight distinct prompt categories. Each sample is meticulously annotated from four evaluation dimensions: semantic alignment, geometric quality, texture quality, and overall quality. The dataset construction process involves generating 160 prompts using large language models, and generating samples via eight prevalent text-to-3D generation methods. MATE-3D aims to address the limitations of existing benchmarks in terms of prompt diversity and multi-dimensional evaluation, providing a comprehensive evaluation tool for research on text-to-3D generation methods.

- 1Benchmarking and Learning Multi-Dimensional Quality Evaluator for Text-to-3D Generation上海交通大学 · 2024年



