Introduction:
The UCF-101 dataset is a widely used benchmark for action recognition in videos. The dataset contains 13,320 videos of 101 action categories, and it was created by collecting YouTube videos and using human annotators to label the action categories. However, the original UCF-101 dataset has certificate issues that may cause difficulties during the download process. Additionally, the dataset is in RAR format, which may not be convenient for some users. Therefore, we have created a ZIP version of the dataset to make it more accessible for researchers and enthusiasts.

Dataset Reproduction:
The dataset was reproduced by downloading the original UCF-101 dataset from the official website (https://www.crcv.ucf.edu/data/UCF101.php) and converting it to ZIP format. The videos were not altered in any way, and the labels and annotations are identical to the original dataset.

Dataset Information:
The UCF-101 dataset consists of 101 action categories, each containing between 24 and 953 videos. The total number of videos is 13,320, and the total size of the dataset is approximately 7.2 GB. The videos have a resolution of 320x240 pixels, and the duration varies between 1 and 30 seconds. The videos were captured from a variety of sources, including YouTube, and they feature people performing different actions, such as playing basketball, riding a bike, or cooking.

Usage:
The UCF-101 dataset can be used for a variety of research projects related to action recognition in videos, such as training and evaluating deep learning models. To use the dataset, simply download the ZIP file from the provided link and extract it to your preferred directory. The dataset is organized by action categories, with each category containing a folder with the corresponding videos. The file names include the action category and the video ID, and the labels are provided in a separate file.

Citation:
If you use this data set, please refer to the following technical report:
Khurram Soomro, Amir Roshan Zamir and Mubarak Shah, UCF101: A Dataset of 101 Human Action Classes From Videos in The Wild., CRCV-TR-12-01, November, 2012.

Contact:
If you have any questions or feedback regarding the UCF-101 dataset reproduction, please contact us at chenyuan.qu@outlook.com