Rule Induction code

Rule induction in the context of automated sports video annotation.

Any questions/comments, please contact Dr. Aftab Khan or Dr. David Windridge.

ACASVA Actions Dataset

Nazli FarajiDavar and Teofilo deCampos


Player's action recognition is one of the challenges in the ACASVA project. The goal is to classify each action sample into three classes: Non-Hit, Hit and Serve. We have done simple classification experiments [3] and transductive transfer learning experiments [1][2]. The splits of data (for training, validation, testing) are done per game match, please see the references below for further detail.

Following deCampos et al [3], we used HOG3D descriptors extracted on player bounding boxes. Two different sets of feature extraction parameters were used: the 960D parameters (4x4x3x20) optimised for the KTH dataset and the 300D parameters (2x2x5x5x3) optimised for the Hollywood dataset (see Alexander Klaser's page for details). In our preliminary experiments, we found that the KTH parameters (960D) give better results for the tennis dataset.

The table below summarises the datasets. Each row refers to a video and a zip file was created for each type of HOG3D vector (dimensionality value). The size of each zip file is from 2 to 4MB.

labelsport gender number competition yearNon-HitHitServedimensionality
TWSA03 TennisWomenSingleAustralia200394421472300 /960[1][3]
TMSA03 TennisMenSingleAustralia20031881469123300 /960 [2]
TWSJ09 TennisMenSingleJapan200985922459300 /960
TWDA09 TennisWomenDoubleAustralia2009106413536300 /960[1][3]
TWDU06 TennisWomenDoubleUS2006326940577300 /960
BMSB08 BadmintonMenSingleBeijing20087064588300 /960[2]

Unfortunately we can not provide the access to the original videos but you can download (above) the extracted features for each action sample in the videos. In each ZIP file, there are four ASCII files, where each line contain information about one sample.

  • labels.txt: contains action labels; Non-Hit (0), Hit (1) and Serve (2);
  • frames.txt: for each sample, it indicates the time stamp of the original video when the features were extracted - note that multiple players are visible in each frame and for this reason consecutive lines have the same frame number;
  • teams.txt: represents players as Far player (0) and Near player (1) where far and near players are decided based on the player's feet position in relative to court's mid-line;
  • features.txt: contains feature vectors which has either 300 or 960 dimensional vectors, extracted using HOG3D - each line represents a feature vector for a action sample. The first element of each line indicates the dimensionality.


    This dataset was designed for the
    transductive transfer learning experiments published in the references below.

    Each file contains HOG3D data extracted from one match of tennis or badminton. Because the conditions are different (illumination, athletes, clothes, etc), this dataset poses an interesting application scenario for transductive transfer learning methods. For example, a game of badminton can be used for training, a first game of tennis can be used for transfer (or domain adaptation) and a second game of tennis can be used as the test set.

    An alternative usage scenario is to evaluate methods that are able to learn the structure of a temporal pattern. The table below shows a typical sequence of actions in a tennis game. The following pattern becomes clear: only one paper can hit the ball at a given instant, and hits are normally followed by inactivity of both players, because when the tennis ball is bouncing on the floor, none of the players are hitting it.

    time stamp action of far player action of near player


    The following papers give further descriptions of this dataset and experiments:

    [1] N. FarajiDavar, T. deCampos, J. Kittler and F.Yan
    Transductive Transfer Learning for Action Recognition in Tennis Games
    In 3rd International Workshop on Video Event Categorization, Tagging and Retrieval for Real-World Applications (VECTaR), in conjunction with 13th Internatinal Conference on Computer Vision, Barcelona, Spain 2011

    [2] N. FarajiDavar, T. deCampos, D. Windridge, J. Kittler and W. Christmas
    Domain Adaptation in the Context of Sport Video Action Recognition
    In Domain Adaptation Workshop, in conjunction with NIPS, Sierra Nevada, Spain 2011

    [3] T. deCampos, M. Barnard, K. Mikolajczyk, J. Kittler, F. Yan, W. Christmas and D. Windridge
    An evaluation of bags-of-words and spatio-temporal shapes for action recognition
    In IEEE Workshop on Applications of Computer Vision (WACV), 2011

    A copy of this dataset is available at the repository, doi:10.5881/ACASVA-ACTIONS-DATASET.


  • Highlights

    January, 2010

    Secure Access

    May, 2009

    Website Developed