Semantic Alignment Benchmark
收藏资源简介:
该数据集是SVG-Score评估框架的核心,由哈索·普拉特纳研究院和摩德纳大学与雷焦艾米利亚大学联合创建,用于评估文本到SVG生成的语义对齐程度。数据集包含超过12,000个由人类标注的caption-SVG对评分,涵盖8,671个SVG和1,858个caption,评分范围为1至5,评估对象包括对象存在性、颜色、数量及空间关系等属性。构建过程从OmniSVG数据集采样caption-SVG对,通过检索扩展并结合人工标注,确保评分的准确性。该数据集旨在解决现有评估指标(如CLIPScore)对SVG语义错误不敏感的问题,为训练和验证人类对齐的SVG评估器提供基准,推动文本到SVG生成模型的可靠评估。
This dataset is the core of the SVG-Score evaluation framework, jointly created by the Hasso Plattner Institute and the University of Modena and Reggio Emilia, for evaluating the semantic alignment of text-to-SVG generation. It contains over 12,000 human-annotated caption-SVG pair scores, covering 8,671 SVGs and 1,858 captions, with scores ranging from 1 to 5. The evaluated attributes include object existence, color, quantity, spatial relationships and other relevant properties. The dataset construction process samples caption-SVG pairs from the OmniSVG dataset, expands them via retrieval and incorporates manual annotation to ensure the accuracy of the scoring. This dataset aims to address the issue that existing evaluation metrics (e.g., CLIPScore) are insensitive to SVG semantic errors, serve as a benchmark for training and validating human-aligned SVG evaluators, and advance reliable evaluation of text-to-SVG generation models.




