UCF50 - Action Recognition Dataset
Click here to check the published results on UCF50 (updated September-12-2012)
UCF50 is an action recognition dataset with 50 action categories, consisting of realistic videos taken from youtube. This dataset is an extension of YouTube Action dataset (UCF11) which has 11 action categories.
Most of the available action recognition datasets are not realistic and are staged by actors. In our dataset, the primary focus is to provide the computer vision community with an action recognition dataset consisting of realistic videos which are taken from youtube. Our dataset is very challenging due to large variations in camera motion, object appearance and pose, object scale, viewpoint, cluttered background, illumination conditions, etc. For all the 50 categories, the videos are grouped into 25 groups, where each group consists of more than 4 action clips. The video clips in the same group may share some common features, such as the same person, similar background, similar viewpoint, and so on.
UCF50 dataset's 50 action categories collected from youtube are: Baseball Pitch, Basketball Shooting, Bench Press, Biking, Biking, Billiards Shot,Breaststroke, Clean and Jerk, Diving, Drumming, Fencing, Golf Swing, Playing Guitar, High Jump, Horse Race, Horse Riding, Hula Hoop, Javelin Throw, Juggling Balls, Jump Rope, Jumping Jack, Kayaking, Lunges, Military Parade, Mixing Batter, Nun chucks, Playing Piano, Pizza Tossing, Pole Vault, Pommel Horse, Pull Ups, Punch, Push Ups, Rock Climbing Indoor, Rope Climbing, Rowing, Salsa Spins, Skate Boarding, Skiing, Skijet, Soccer Juggling, Swing, Playing Tabla, TaiChi, Tennis Swing, Trampoline Jumping, Playing Violin, Volleyball Spiking, Walking with a dog, and Yo Yo.
The dataset can be downloaded by clicking here
If you use this dataset, please refer to the following paper:
Kishore K. Reddy, and Mubarak Shah, "Recognizing 50 Human Action Categories of Web Videos", Machine Vision and Applications Journal (MVAP), September 2012.
For questions regarding this dataset, please contact Kishore Reddy.
Results on UCF50
If you happen to use UCF50, send us an email with the following details and we will update our webpage with your results- Performance (%)
- Experimental Setup (In order to keep the reported results consistant, please follow "Leave One Group Out Cross Validation" which will lead to 25 cross-validations. This would eliminate randomness in the experimental setup)
- Paper details
| Performance | Experimental Setup | Paper |
|---|---|---|
| 76.90% | Leave One Group Out Cross-validation (25 cross-validations) | Reddy and Shah. (MVAP), 2012 |
| 57.90% | Group Wise Cross-validation | Sadanand and Corso. (CVPR), 2012 |
| 76.40%* | Video Wise Cross-validation (*Since videos belonging to a group are obtained from a single long video, similar videos can end up in both training and testing in "video-wise cross-validation" leading to high performance) | Sadanand and Corso. (CVPR), 2012 |
| 81.03%* | 2/3 training and 1/3 testing for each class (*From the details given in the paper, we are not sure if videos belonging to the same group are kept seperate in training and testing sets and the paper does not give details on number of cross-validations) | Todorovic. (ECCV), 2012 |
| 73.70% | Leave One Group Out Cross-validation (25 cross-validations) | Solmaz, et al. (MVAP), 2012 |