LEMMA
收藏资源简介:
LEMMA数据集由加州大学洛杉矶分校视觉、认知、学习与自主性中心创建,专注于多代理多任务日常活动的学习。数据集包含324个活动,涵盖单代理单任务、单代理多任务、多代理单任务和多代理多任务四种场景。通过精心设计的设置,LEMMA数据集密集标注了原子动作与人类-物体交互,以提供日常活动的构成性、调度性和分配性的基准真相。此外,LEMMA还设计了复合动作识别和动作/任务预测基准,以衡量对复杂目标导向活动的理解能力和时间推理能力。该数据集旨在推动机器视觉社区对目标导向人类活动的研究,并进一步探索现实世界中的任务调度和分配。
The LEMMA dataset was created by the Center for Vision, Cognition, Learning and Autonomy at the University of California, Los Angeles (UCLA), focusing on the learning of multi-agent and multi-task daily activities. The dataset comprises 324 activities, covering four scenarios: single-agent single-task, single-agent multi-task, multi-agent single-task, and multi-agent multi-task. Through carefully designed experimental settings, the LEMMA dataset densely annotates atomic actions and human-object interactions, providing ground truth for the compositional, scheduling, and distributive characteristics of daily activities. Additionally, the LEMMA dataset also establishes benchmarks for composite action recognition and action/task prediction, aiming to evaluate the ability to understand complex goal-directed activities and perform temporal reasoning. This dataset is intended to advance research on goal-directed human activities in the computer vision community, and further explore task scheduling and distribution in real-world scenarios.




