A correctly recognized formation
The strategic design of a football play begins with the initial formation. Thus, the first part of our analysis of each play is the recognition of the initial formation. Recognized formations are also used in later stages of our system, for example to initialize player trackers.
Formation recognition entails determining both the types and the locations of the players on the field. This is an interesting and difficult problem because the types of players in a formation can vary from one to another, and the locations of the players in the formation depend on the players’ types. Moreover, there are several constraints imposed by the rules of football that limit what types of players can be in a formation together and what their locations can be.
These aspects of the problem make it difficult to use current object recognition methods directly on football formations. To cope with these difficulties, we propose the mixture-of-parts pictorial structure model (MoPPS), which is an extension of classical pictorial structures that permits reasoning about objects whose parts can vary (here we think of a football formation as an object and of the players as its parts). The papers below describe the MoPPS model in more detail.
Dataset
Our football formation dataset is available for public use. See this page for more details.
Papers
- R. Hess, A. Fern and E. Mortensen. Mixture-of-parts pictorial structures for objects with variable part sets. In Proc. IEEE International Conf. on Computer Vision, 2007.
- R. Hess and A. Fern. Toward learning mixture-of-parts pictorial structures. In the ICML 2007 Workshop on Constrained Optimization and Structured Output Spaces, 2007.