数据集

we collect and construct a large scale conducting motion dataset. Based on advanced object detection and pose estimation techniques, we efficiently extract conducting motion data from online video sources. The constructed dataset, named ConductorMotion100, has 100 hours of conducting motion data and aligned Mel spectrograms. Its scale significantly exceeds that of existing conducting motion datasets.

we collect a video dataset of 17 drivers containing fatigue state or non-fatigue state, where the videos of fatigue state is collected when the drivers have sufficient sleeping, while non-fatigue state videos are collected under sleep deprivation condition. Each video lasts about 100 seconds. We use a 300-frame sliding-window and a stride of 5 frames to collect training and testing samples. Each sample consists of 300 video frames and a label indicating the corresponding fatigue state. The samples are divided into a training set and a testing set whose sizes are respectively 13,402 and 7,332.

We construct a large-scale Multi-modal E-commerce Products classification dataset MEP-3M, which consists of over 3 million products and 599 fine-grained product categories. Each product is represented with an image-text pair and annotated with hierarchical labels.