InstructIAD
收藏资源简介:
InstructIAD数据集是一个针对工业异常检测的指令微调数据集,由哈尔滨工业大学创建。该数据集包含9444个异常样本和13578个正常样本,涵盖40种不同的产品类型。数据集中的每个样本都带有丰富属性的详细描述,描述产品的颜色、形状、布局、材料和纹理等特征,以及缺陷的位置、方向、形状和颜色。InstructIAD支持三种关键任务:异常检测、属性级描述和异常分析,旨在帮助模型更好地理解视觉属性和制造流程,以提高工业异常检测的准确性。
The InstructIAD dataset is an instruction-tuning dataset for industrial anomaly detection, created by Harbin Institute of Technology. It contains 9,444 anomalous samples and 13,578 normal samples, covering 40 distinct product categories. Each sample in the dataset comes with detailed descriptions covering rich attributes, including product features such as color, shape, layout, material and texture, as well as the location, orientation, shape and color of defects. The InstructIAD dataset supports three core tasks: anomaly detection, attribute-level description and anomaly analysis, aiming to help models better understand visual attributes and manufacturing processes to improve the accuracy of industrial anomaly detection.

- 1Triad: Empowering LMM-based Anomaly Detection with Vision Expert-guided Visual Tokenizer and Manufacturing Process哈尔滨工业大学 · 2025年



