SteelBench
收藏资源简介:
SteelBench是由印度理工学院比莱分校与印度钢铁管理局联合创建的工业监控诊断基准数据集,旨在评估视觉语言模型在真实工业环境中的性能。该数据集包含1345个密集标注的视频剪辑,源自149小时的真实钢铁厂监控录像,涵盖25个动作类别、个人防护装备评估及安全合规性标注,总计超过42000个标注决策。其构建过程采用了时间去重、分层采样及四阶段标注流程,并引入了来源感知审计协议以确保标签可靠性。该数据集主要应用于工业安全监控领域,旨在解决模型在复杂视觉条件(如粉尘、蒸汽、低光照)下的活动识别、安全规则推理及标注来源依赖性等核心问题。
GEO-Bench-2 is a comprehensive geospatial artificial intelligence evaluation benchmark jointly constructed by multiple institutions, covering 19 rigorously curated open-licensed datasets. This benchmark integrates five major task types including classification, segmentation, detection and others. Its data sources span all seven continents across the globe, and include multi-modal data such as multispectral imagery, SAR data and time-series data, adopting the TACO standard format to ensure machine learning readiness. The dataset is constructed via geographic balanced sampling and capability grouping strategies, aiming to systematically evaluate the generalization capabilities of geospatial foundation models in practical scenarios such as agricultural monitoring and disaster response, and promote the development of general geospatial intelligence technologies.

- 1GEO-Bench-2: From Performance to Capability, Rethinking Evaluation in Geospatial AIIBM研究院欧洲分部, 慕尼黑工业大学, 克拉克大学, MBZUAI, 美国宇航局, 欧洲航天局, 亚利桑那州立大学, ServiceNow研究院 · 2025年



