ECBench
收藏资源简介:
ECBench是由阿里巴巴达摩院、浙江大学和同济大学联合开发的一个多模态基准测试数据集,旨在评估大型视觉语言模型在自我中心视频中的认知能力。该数据集包含386个RGB-D视频和4324个问答对,涵盖了30个不同的认知维度,包括感知、推理、自我意识、动态捕捉和幻觉等。数据集的创建过程采用了类独立的人工标注和多轮问题筛选策略,确保了数据的质量和平衡性。ECBench的应用领域主要集中在机器人技术和人工智能领域,旨在解决机器人在动态环境中进行复杂任务时的认知挑战。
ECBench is a multimodal benchmark dataset jointly developed by Alibaba DAMO Academy, Zhejiang University and Tongji University, aiming to evaluate the cognitive capabilities of large vision-language models in egocentric videos. This dataset contains 386 RGB-D videos and 4324 question-answer pairs, covering 30 distinct cognitive dimensions including perception, reasoning, self-awareness, dynamic capture and hallucination. The dataset was constructed using class-independent manual annotation and multi-round question screening strategies to ensure data quality and balance. The application scenarios of ECBench mainly focus on robotics and artificial intelligence, and it is designed to address the cognitive challenges faced by robots when performing complex tasks in dynamic environments.

- 1ECBench: Can Multi-modal Foundation Models Understand the Egocentric World? A Holistic Embodied Cognition Benchmark阿里巴巴达摩院, 浙江大学, 同济大学 · 2025年



