GlazyBench
收藏资源简介:
GLAZY Benchmark是一个用于陶瓷釉料属性预测和条件图像生成的开源基准测试数据集。该数据集包含两个独立的基准测试轨道:属性预测和图像生成。属性预测轨道包含四个具体任务:透明度分类(4个类别:不透明、半不透明、半透明、透明)、表面质地分类(9个类别:光泽、半光泽、缎面、缎面哑光、哑光、半哑光、光滑哑光、干燥哑光、石质哑光)、颜色族分类(9个类别:黑、蓝、灰、绿、橙、紫、红、白、黄)以及RGB颜色回归。图像生成轨道包含釉料图像及其对应的结构化元数据,用于条件图像生成任务。数据集采用固定的训练/测试划分:属性预测轨道包含16,781个训练样本和4,903个测试样本;图像生成轨道包含4,490个训练样本和443个测试样本。每个任务在训练集和测试集中都有明确的标注覆盖数量统计。数据集还提供了可选的源信息层,包括原始HTML页面、解析后的元数据以及工具脚本,支持用户进行自定义提取。此外,数据集包包含了轻量级基线脚本,包括属性预测的多数类基线和RGB均值基线,以及图像生成的最近邻检索基线。该数据集专为机器学习研究设计,适用于多分类、回归和条件图像生成等任务的基准测试和模型评估。
GLAZY Benchmark is an open-source benchmark dataset for ceramic glaze property prediction and conditional image generation. The dataset includes two independent benchmark tracks: property prediction and image generation. The property prediction track comprises four specific tasks: transparency classification (4 categories: opaque, semi-opaque, semi-transparent, transparent), surface texture classification (9 categories: glossy, semi-glossy, satin, satin matte, matte, semi-matte, smooth matte, dry matte, stone matte), color family classification (9 categories: black, blue, gray, green, orange, purple, red, white, yellow), and RGB color regression. The image generation track includes glaze images and their corresponding structured metadata for conditional image generation tasks. The dataset employs fixed train/test splits: the property prediction track contains 16,781 training samples and 4,903 test samples; the image generation track contains 4,490 training samples and 443 test samples. Each task has explicit annotation coverage statistics in both the training and test sets. The dataset also provides optional source information layers, including raw HTML pages, parsed metadata, and tool scripts, supporting user-defined extraction. Furthermore, the dataset package includes lightweight baseline scripts, such as majority class baseline and RGB mean baseline for property prediction, and nearest neighbor retrieval baseline for image generation. This dataset is designed for machine learning research and is suitable for benchmark testing and model evaluation in tasks like multi-class classification, regression, and conditional image generation.




