COCO Person FaceSwap (COCO-PFS)
收藏资源简介:
COCO Person FaceSwap (COCO-PFS) 数据集由意大利国家研究委员会信息科学与技术研究所开发,旨在为身份感知跨模态检索任务提供大规模的训练和评估数据。该数据集基于广泛使用的COCO数据集,通过深度伪造技术将其中的人脸替换为VGGFace2中的公共人物面孔,并生成了包含500个不同实体的49,957张图像。数据集的内容包括图像及其对应的描述,描述中明确提到了替换后的人物姓名,以支持身份感知检索任务。数据集的创建过程包括图像预选、人脸替换和描述增强等步骤,确保了数据的多样性和实用性。该数据集的应用领域主要集中在个性化视频检索、大规模音视频档案管理以及文化传承等领域,旨在解决现有跨模态检索模型在处理特定人物身份和上下文信息时的局限性。
COCO Person FaceSwap (COCO-PFS) dataset was developed by the Institute of Information Science and Technologies of the Italian National Research Council, with the goal of providing large-scale training and evaluation data for identity-aware cross-modal retrieval tasks. Built upon the widely adopted COCO dataset, this dataset generates 49,957 images encompassing 500 distinct entities by replacing human faces in the original COCO dataset with public person faces from the VGGFace2 dataset via deepfake technology. The dataset includes images and their corresponding captions, which explicitly mention the names of the swapped individuals to support identity-aware retrieval tasks. The development pipeline of the dataset comprises image pre-selection, face swapping, and caption enhancement, ensuring the diversity and practicality of the data. Its main application scenarios include personalized video retrieval, large-scale audio-visual archive management, and cultural heritage preservation, aiming to address the limitations of existing cross-modal retrieval models when handling specific person identities and contextual information.




