Egoshots

Egoshots数据集是由法国国家信息与自动化研究所创建的，包含978张真实生活场景的图像，旨在评估图像字幕模型的多样性和鲁棒性。该数据集通过Autographer相机随机拍摄，涵盖室内外多种场景，用于解决现有图像字幕模型在日常生活中的应用限制问题。数据集的创建过程涉及使用预训练的图像字幕和对象识别网络进行图像标注，并提出了一种新的图像字幕评估指标——基于对象的语义保真度（SF），以评估无标注图像的字幕质量。Egoshots数据集的应用领域包括支持视觉障碍人士和推荐系统等，旨在提高这些系统的可靠性和实用性。

The Egoshots dataset was developed by the Institut National de Recherche en Informatique et en Automatique (INRIA), consisting of 978 images depicting real-life scenarios. Its core objective is to evaluate the diversity and robustness of image captioning models. Captured randomly via Autographer cameras, the dataset covers a wide range of indoor and outdoor scenes, aiming to address the application limitations of existing image captioning models in daily life. The construction of this dataset involves using pre-trained image captioning and object recognition networks for image annotation, and proposes a novel evaluation metric for image captioning: Object-based Semantic Fidelity (SF), which is used to assess the caption quality of unannotated images. Application scenarios of the Egoshots dataset include assisting visually impaired individuals and recommendation systems, with the goal of improving the reliability and practicality of these systems.