R3

This dataset contains the visual features of over 1500 days of lifelogging, recorded by more than 50 independent subjects*. Due to privacy constraints, the original pictures are not provided. We release the dataset in two independent subsets in case it is to be used in conjunction with our Visual Context Predictor model. Such model is trained with data from R3training.h5, and has not seen any data from R3testing.h5.

Each of the data files contains four hdf5 datasets: 'user_id', 'day' (consecutive recording days), 'frame_id' (HHMMSS), and 'descriptor' (the visual feature for that frame). The features are unnormalized, as extracted using the Keras model InceptionV3 with parameters include_top=False, pooling='max'. Note that the VCP model has been trained with data normalized with min-max normalization to be between 0 and 1.

If you use this data or the model, please cite the following publication:

Ana García del Molino, Joo-Hwee Lim, and Ah-Hwee Tan. 2018. Predicting Visual Context for Unsupervised Event Segmentation in Continuous Photo-streams. In Proceedings of ACM Multimedia conference (ACMMM’18). ACM,
New York, NY, USA. https://doi.org/10.1145/3240508.3240624

(*) Due to storage space limitations the dataset had to be slightly reduced compared to the numbers reported on our paper "Predicting Visual Context for Unsupervised Event Segmentation in Continuous Photo-streams.". If you require other visual features for your research, please contact us.

本数据集汇聚了超过1500天的生活日志视觉特征，由50位独立受试者所记录。鉴于隐私保护限制，原始图片并未提供。为便于与我们的视觉情境预测模型相结合使用，本数据集被划分为两个独立的子集。所述视觉情境预测模型系采用R3training.h5数据集进行训练，并未接触过R3testing.h5数据集中的任何信息。每个数据文件包含四个hdf5数据集：'user_id'（用户ID），'day'（连续记录日），'frame_id'（时分秒帧ID），以及'descriptor'（该帧的视觉特征）。特征值未经标准化处理，系采用Keras模型InceptionV3（参数包括top层不包含、池化方式为最大池化）提取。请注意，视觉情境预测模型系采用最小-最大标准化方法，将数据归一化至0至1之间进行训练。若您使用此数据或模型，请引用以下出版物：Ana García del Molino, Joo-Hwee Lim, 和 Ah-Hwee Tan. 2018. 在连续照片流中预测视觉情境以实现无监督事件分割。发表于ACM多媒体会议（ACMMM’18）论文集。ACM，纽约，纽约州，美国。https://doi.org/10.1145/3240508.3240624。（*）由于存储空间限制，数据集的规模略小于我们在《在连续照片流中预测视觉情境以实现无监督事件分割》一文中报告的规模。如需其他视觉特征用于研究，请与我们联系。