UCF Sports Action Data Set
Khurram Soomro and Amir R. Zamir, Action Recognition in Realistic Sports Videos, Computer Vision in Sports, Springer International Publishing, 2014.
UCF Sports dataset consists of a set of actions collected from various sports which are typically featured on broadcast television channels such as the BBC and ESPN. The video sequences were obtained from a wide range of stock footage websites including BBC Motion gallery and GettyImages.
The dataset includes a total of 150 sequences with the resolution of 720 x 480. The collection represents a natural pool of actions featured in a wide range of scenes and viewpoints. By releasing the data set we hope to encourage further research into this class of action recognition in unconstrained environments. Since its introduction, the dataset has been used for numerous applications such as: action recognition, action localization, and saliency detection.
Data Set Details
The dataset includes the following 10 actions. The figure above shows the a sample frame of all ten actions, along with their bounding box annotations of the humans shown in yellow.
- Diving (14 videos)
- Golf Swing (18 videos)
- Kicking (20 videos)
- Lifting (6 videos)
- Riding Horse (12 videos)
- Running (13 videos)
- Skate-Boarding (12 videos)
- Swing-Bench (20 videos)
- Swing-Side (13 videos)
- Walking (22 videos)
The following table summarizes the characteristics of the data-set:
It is evident that certain actions are short in nature, such as kicking, as compared to walking or running, which are relatively longer and have more periodicity. However, it is apparent from the chart that the average duration of action clips shows great similarities across different classes. Therefore, merely considering the duration of one clip would not be enough for identifying the action.
- Action Recognition [Leave-One-Out (LOO) cross-validation scheme]: It is recommended as in  to test on UCF Sports by using a Leave-One-Out (LOO) cross-validation scheme. This scenario takes out one sample video for testing and trains using all of the remaining videos of an action class. This is performed for every sample video in a cyclic manner, and the overall accuracy is obtained by averaging the accuracy of all iterations.
- Action Localization [Train/Test Splits]: It is recommended to use train/test splits as suggested in [*]. The proposed experimental setup splits the dataset into two uneven parts: two-third of videos for training and one-third for testing. To calculate the accuracy, an intersection-over-union criterion is used to plot ROC curves with a certain overlap threshold. The intersection-over-union computes the overlap between the predicted bounding box and the ground truth, and divides it by the union of both the bounding boxes, for every frame. This value is then averaged over all frames in a video. A 20 % overlap threshold is used for this experiment. Area Under Curve (AUC) against the overlap threshold, which shows how the performance varies if the threshold is changed, is used to compute the final performance. To calculate the overlap, the ground truth bounding box per frame is provided for the dataset.
[*] Tian Lan, Yang Wang and Greg Mori, Discriminative figure-centric models for joint action localization and recognition, IEEE International Conference on Computer Vision (ICCV), 2011.
Mikel D. Rodriguez, Javed Ahmed, and Mubarak Shah, Action MACH: A Spatio-temporal Maximum Average Correlation Height Filter for Action Recognition, Computer Vision and Pattern Recognition, 2008.