 |
 |
 |
A frame of football video before and after registration with a 2D model of the football field
A necessary first step for many components of our system is to determine exactly what location on the field is depicted in a particular frame of video. For example, to classify the initial formation of the offense, we would like to know exactly where on the field the players are standing. This knowledge, in turn, allows us to define models describing the ideal positions of players relative to other players on the field.
The easiest way to determine a video frame’s location on the field is to register the frame with a static 2D model of the football field, as depicted above. One common approach to video registration is to match distinctive image features, such as SIFT keypoints, to a pre-assembled model. Unfortunately, in football video, many frames lack distinctive image features and thus cannot be robustly registered using this approach.
With this in mind, we developed a registration method that overcomes this difficulty using a concept of local distinctiveness to find matches between non-distinctive features, such as the hash marks on the football field which can be found in nearly every frame of our video. For further details, refer to our paper, which describes our method more completely and presents experimental results demonstrating its effectiveness. You may also be interested in our registration datset.
Video Results
Below are links to some video results from our registration method.
Each clip shows the original input video on top and the registration
results below that.
- register_01.mpeg
- register_02.mpeg
- register_03.mpeg
- register_04.mpeg
- register_05.mpeg
- register_06.mpeg
- register_07.mpeg
- register_08.mpeg
Dataset
Our registration dataset from the football domain is available. See this page for more details.
Papers
- R. Hess and A. Fern. Improved video registration using non-distinctive local image features. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2007.