wmbench
收藏资源简介:
WMBench 是一个用于评估世界模型(如 Cosmos、OpenSora 等)生成的视频中是否存在幻觉或异常的基准测试套件,其评估标准基于真实的机器人训练数据分布。该数据集旨在为视频生成模型的幻觉检测和异常检测任务提供评估基准。数据集当前包含一个主要部分:`gr-1/`,其视频数据来源于 NVIDIA 的 GR1 机器人项目(具体为 PhysicalAI-Robotics-GR00T-GR1)。该部分包含 5 个评估任务,共计提供了 5 个真实机器人训练视频和 24 个由 Cosmos 世界模型生成的视频。此外,计划未来纳入来自斯坦福大学 DROID 项目的 `droid/` 数据集。
WMBench is a benchmark test suite for evaluating hallucinations or anomalies in videos generated by world models (such as Cosmos, OpenSora, etc.), with evaluation criteria based on real robot training data distributions. This dataset aims to provide an evaluation benchmark for hallucination detection and anomaly detection tasks in video generation models. The dataset currently includes a main section: `gr-1/`, with video data sourced from NVIDIAs GR1 robot project (specifically PhysicalAI-Robotics-GR00T-GR1). This section contains 5 evaluation tasks, providing a total of 5 real robot training videos and 24 videos generated by the Cosmos world model. Additionally, there are plans to incorporate the `droid/` dataset from Stanford Universitys DROID project in the future.





