Football formation detection research wins best paper award
ADSC researchers were awarded the best paper award at the first IEEE International Workshop on Computer Vision in Sports (CVSports), which was held in conjunction with the IEEE Computer Vision and Pattern Recognition (CVPR) conference, in June.
Using over 800 American football play clips, including test video from the University of Illinois football team, ADSC researchers proposed a novel framework where they are able to automatically identify the line of scrimmage, the formation of the offensive team for that play and the formation frame, which is the time instant right before the ball is snapped and the play is started, in a football video.Narendra Ahuja, the research was conducted by ADSC researchers Indriyati Atmosukarto and Shaunak Ahuja, Adjunct Senior Research Scientist Bernard Ghanem and former ADSC intern Karthik Muthuswamy.
Their proposed method is 95 percent accurate in detecting the formation frame, 98 percent accurate in detecting the line of scrimmage and up to 67 percent accurate in classifying the offensive team’s formation.
In its inaugural workshop, CVSports aims to bring together practitioners and researchers from different disciplines to share ideas and methods on current and future use of computer vision in sports, as computer vision has begun to play an important role in sports, such as real-time enhanced viewing, better understanding of sports injuries and automatic annotation of video footage.
“It is the ideal venue for our work as it focused on the new ways computer algorithms and methods are applied to sports,” Atmosukarto said.
While automatic action analysis is commonplace in areas such as security surveillance and military applications, most sports teams use manual annotation when analyzing sports videos. This process is extremely time consuming and the repetitive nature of the annotation often leads to errors.
Automatically detecting the line of scrimmage or an offensive formation is difficult for a computer due to similar appearances in players or formations, as well as players not being visible in portions of play clips. However, the ability to detect these two formations is a fundamental building block to any future work on action recognition and sports play understanding.
ADSC researchers look forward to using this automatic formation classification framework to enhance their sports video analysis prototype, AutoScout, to automatically identify personnel and to better understand plays and the strategic playbook inference of a team.
“The framework we’ve developed here is yet another building block of AutoScout for automatic analysis of American football videos, as well as videos of other sports in the future,” Atmosukarto said.