This tarball contains the annotations for the UC Irvine's Men's Basketball
dataset, introduced in our ECCV 2010 paper. The annotations may have some rough
edges, but they are mostly correct. These annotations were last updated on
July 14, 2011.

The scenes.txt file contains the scenes that we deemed useful for computer
vision research. In this file, each line is an scene. The first number is the
start frame and the last number is the end frame. Frames outside of this range
may not contain useful basketball footage (e.g., sportscaster commentary).

The people.txt and ball.txt files contain the annotations in text format. Each
row is an annotation. The columns are defined as:

    1   Path ID     rows with the same ID are the same object
    2   xmin / xtl  top left x-coordinate of the bounding box
    3   ymin / ytl  top left y-coordinate of the bounding box
    4   xmax / xbr  bottom right x-coordinate of the bounding box
    5   ymax / ybr  bottom right y-coordinate of the bounding box
    6   frame       frame number of the annotation
    7   lost?       if true, then the object is outside of the frame
    8   occluded?   if true, the object is inside the frame, but occluded
    9   generated?  if true, this annotation was interpolated
    10  class       the class label for this object

The people.mat and ball.mat file contains the annotations in MATLAB format,
following the same conventions as above.

The corresponding frames are available online:
    http://phoenix.ics.uci.edu/cvondrick/basketball.tar.gz (12.7 GB)

If you use this data set, please cite:

    Carl Vondrick, Deva Ramanan, Donald Patterson. "Efficiently Scaling Up
    Video Annotation with Crowdsourced Marketplaces" European Conference on
    Computer Vision (ECCV) Crete, Greece, September 2010.

Contact Carl Vondrick (cvondric@ics.uci.edu) with any questions.
