Gaze360 (Physically Unconstrained Gaze Estimation in the Wild)

了解人们在看哪里是一个信息丰富的社交线索。在这项工作中，我们介绍了Gaze360，这是一种大规模的凝视跟踪数据集和方法，用于在不受约束的图像中进行稳健的3D凝视估计。我们的数据集由室内和室外环境中的238个主题组成，带有标记的3D凝视，跨越各种头部姿势和距离。通过简单有效的收集方法，它是同类主题和多样性中最大的公开数据集。我们提出的3D凝视模型扩展了现有模型，以包括时间信息并直接输出凝视不确定性的估计值。我们通过消融研究证明了我们模型的优势，并通过与其他近期凝视基准数据集的跨数据集评估显示了其泛化性能。我们还提出了一种简单的自我监督方法来改善跨数据集域适应。最后，我们演示了我们的模型在超市环境中估计客户注意力的应用。

Knowing where people are looking is an informative social cue. In this work, we introduce Gaze360, a large-scale gaze tracking dataset and method for robust 3D gaze estimation in unconstrained images. Our dataset comprises 238 subjects in both indoor and outdoor environments, with annotated 3D gaze labels, covering a wide range of head poses and distances. Through a simple yet effective collection pipeline, it is the largest public dataset of its kind in terms of subject count and diversity. The proposed 3D gaze model extends existing models by incorporating temporal information and directly outputting estimates of gaze uncertainty. We validate the advantages of our model through ablation studies, and demonstrate its generalization performance via cross-dataset evaluations against other recent gaze benchmark datasets. We further propose a simple self-supervised method to improve cross-dataset domain adaptation. Finally, we demonstrate the application of our model for estimating customer attention in supermarket environments.

搜集汇总

背景与挑战

背景概述

Gaze360是一个大规模3D凝视估计数据集，包含238个主题在室内外环境下的多样化凝视数据，支持头部姿势和距离变化。该数据集还提供了扩展的3D凝视模型，能估计凝视不确定性并应用于实际场景如超市客户注意力分析。