OVDEval
收藏资源简介:
OVDEval是由西北工业大学等机构联合创建的综合性开放词汇检测评估数据集,包含20000张高质图像和3000个精细标注的标签。该数据集分为9个子集,涵盖6个语言方面,旨在全面测试模型在常识知识、属性理解、位置理解、对象关系理解等方面的能力。通过精心设计,确保所有负标签均为挑战性强的硬负样本,从而严格测试模型的真实理解能力。OVDEval不仅揭示了现有开放词汇检测模型的弱点,还指导了未来研究的方向,特别是在处理复杂语言描述的检测任务上。
OVDEval is a comprehensive open-vocabulary detection and evaluation dataset jointly created by Northwestern Polytechnical University and other institutions. It comprises 20,000 high-quality images and 3,000 meticulously annotated labels. The dataset is divided into 9 subsets covering 6 linguistic aspects, aiming to comprehensively evaluate the capabilities of models in common-sense knowledge, attribute understanding, location understanding and object relational understanding. Through deliberate design, all negative labels are ensured to be challenging hard negatives, so as to rigorously test the actual understanding capabilities of models. OVDEval not only reveals the weaknesses of existing open-vocabulary detection models, but also guides future research directions, particularly in detection tasks involving complex linguistic descriptions.




