WorldLens-26K
收藏资源简介:
WorldLens-26K是由WorldBench团队构建的大规模标注数据集,旨在全面评估生成式世界模型的综合性能。该数据集包含26,000条人类标注的视频样本,每条样本均附有定量评分和文本解释,涵盖视觉真实性、几何一致性、物理合理性和行为安全性等多维指标。数据通过专业标注流程收集,结合了算法生成视频与人工理性标注,为世界模型的真实性评估提供了标准化基准。其核心应用领域为自动驾驶和具身智能,通过量化生成世界的物理与行为可靠性,推动仿真环境从‘视觉逼真’向‘行为合理’的范式升级。
WorldLens-26K is a large-scale annotated dataset constructed by the WorldBench team, aiming to comprehensively evaluate the overall performance of generative world models. This dataset includes 26,000 human-annotated video samples, each paired with quantitative scores and textual explanations, covering multiple evaluation metrics such as visual authenticity, geometric consistency, physical plausibility, and behavioral safety. The dataset is collected through a professional annotation pipeline, integrating algorithm-generated videos and human expert annotations, thereby providing a standardized benchmark for the authenticity evaluation of generative world models. Its core application scenarios are autonomous driving and embodied intelligence. By quantifying the physical and behavioral reliability of the generated world, it promotes the paradigm shift of simulation environments from "visually realistic" to "behaviorally plausible".

- 1WorldLens: Full-Spectrum Evaluations of Driving World Models in Real WorldWorldBench团队 · 2025年



