Context The dataset was created to train convolutional neural networks to locate balls and hands in live video of juggling. For your convenience three pretrained models have also been uploaded. On https://github.com/rasmusakerlund/juggling-vision-py/ you can find the code use the models or train your own. Content The dataset consists of 180 videos of 300 frames each. In data/frames the 54000 frames can be found as 256x256 images. In data/annotations, 180 csv-files are located. Each csv-files consists of 300 lines. Each line starts with the name of the corresponding image file in the data/frames folder and continues with the x and y coordinates for the right hand, the left hand and the balls. Which videos belong to the train, validation, and testset is indicated by the files trainvideos, validationvideos and testvideos. A video of the content of the dataset is available here: https://www.youtube.com/watch?v=ImvfLaX2GbQ Acknowledgements The dataset wouldn't be what it is if it wasn't for all the jugglers that shared their time and their props. A big thank you to all of you!