ManiGaze
收藏资源简介：
Description Current systems for gaze estimation are usually trained and evaluated on datasets with relatively ideal near frontal head poses and visual targets in front of the user. While this make it useful for model design and method comparison, their performance in other realistic sensing conditions and setups is relatively unknown. Developing datasets for such situations is thus needed, both to evaluate robustness of above methods, measure their performance, understand their limitations, and trigger new research to push further the state of the art of gaze tracking. The ManiGaze dataset was designed with these goals in mind. More specifically, it was created to evaluate gaze estimation from remote RGB and RGB-D (standard vision and depth) sensors in Human-Robot Interaction (HRI) settings, and more specifically during object manipulation tasks. The recording methodology was designed to let the user behave freely and encourage a natural interaction with the robot, as well as to automatically collect gaze targets, since a-posteriori annotation is almost impossible for gaze. The dataset involves 17 person who performed four different tasks in four sessions: Marker on the table Targets (MT) session. The robot asks the user to look or point at markers located on a table placed between the robot and himself. End-effector Targets (ET) session. The robot asks the user to look at its end-effector as it moves them in the space between them. Object Manipulation (OM) session. The robot asks the user to perform a sequence of pick-and-place actions using different objects. Set the Table (ST) session. The user is asked to show and explain to the robot how to set a table, with plate, knife and fork, spoon, and glass. The gaze ground truth was automatically recorded for the first two session, providing a convenient benchmark to evaluate gaze estimation methods. The two last sessions provide additional material for further research (e.g. eye-hand coordination, movement analysis, ...). Reference If you use this dataset, please cite the following publication: R. Siegfried, B. Aminian and J.-M. Odobez. ManiGaze: a Dataset for Evaluating Remote Gaze Estimator in Object Manipulation Situations. In Symposium on Eye Tracking Research and Applications, June, 2020.
当前的注视估计（Gaze Estimation）系统通常在头部姿态接近正脸、视觉目标位于用户正前方的相对理想数据集上进行训练与评估。尽管这类数据集有助于模型设计与方法对比，但现有系统在其他真实感知场景与配置下的性能仍鲜为人知。因此，针对此类场景构建数据集十分必要，既用于评估上述方法的鲁棒性、量化其性能、明晰其局限性，也能推动相关研究进一步提升注视追踪（Gaze Tracking）领域的技术前沿水平。 ManiGaze数据集正是基于上述目标开发而成。具体而言，该数据集旨在评估人机交互（Human-Robot Interaction, HRI）场景下，基于远程RGB、红-绿-蓝-深度（RGB-D，即标准视觉与深度）传感器的注视估计性能，尤其聚焦于物体操作任务场景。 数据采集流程的设计允许用户自由行动，鼓励其与机器人开展自然交互，同时可自动采集注视目标——由于注视数据的事后标注几乎无法实现。该数据集共包含17名受试者，他们在四个采集阶段中完成四类不同任务： 桌面标记目标（Marker on the table Targets, MT）阶段：机器人会要求用户注视或指向放置于自身与机器人之间桌面上的标记物。 末端执行器目标（End-effector Targets, ET）阶段：机器人会在自身与用户之间的空间内移动末端执行器，并要求用户注视该执行器。 物体操作（Object Manipulation, OM）阶段：机器人会要求用户使用各类物品完成一系列取放操作。 餐桌布置（Set the Table, ST）阶段：用户需向机器人演示并讲解如何使用餐盘、刀叉、勺子与玻璃杯完成餐桌布置。 前两个阶段的注视真值可自动采集，为注视估计方法的评估提供了便捷的基准测试集。后两个阶段则为进一步研究（如眼手协调、运动分析等）提供了额外的研究素材。 引用说明：若使用该数据集，请引用以下文献：R. Siegfried、B. Aminian与J.-M. Odobez，《ManiGaze：面向物体操作场景的远程注视估计器评估数据集》，发表于2020年6月的眼追踪研究与应用研讨会（Symposium on Eye Tracking Research and Applications）。
