September 13, 2021 feature
When reporting connected sports games unrecorded oregon remotely, commentators should beryllium capable to rapidly admit the numbers connected the players' jersey shirts, arsenic this allows them to support up with what's happening and pass it to their audience. However, rapidly identifying players successful sports videos is not ever easy, arsenic these videos are often taken astatine a region to seizure the wide progression of the game. A further trouble is the accelerated question of the broadcast camera that often results successful question blur.
Researchers astatine University of Waterloo person precocious developed a machine-learning method that tin automatically admit jersey numbers of players successful images extracted from broadcast sports videos. This technique, presented successful a insubstantial pre-published connected arXiv, could assistance to place the jersey numbers of squad players during sports events faster and much efficiently than different existing computational methods.
"Sports jersey fig designation networks successful existing literature consider jersey fig designation arsenic a classification occupation and either (1) see the jersey numbers arsenic abstracted classes (holistic representation), oregon (2) dainty the 2 digits successful a jersey fig arsenic 2 autarkic classes (digit-wise representation)," Kanav Vats, 1 of the researchers who carried retired the study, told Tech Xplore. "For example, the jersey fig '12' tin beryllium modeled by considering '12' arsenic a abstracted people and besides by splitting the fig '12' into 2 constituent digits '1' and '2' and treating the 2 digits arsenic abstracted classes."
Past studies person recovered that learning aggregate output representations tin amended the show of heavy neural networks. In different words, neural networks that are trained to absorption connected antithetic aspects of the task they are learning to implicit were recovered to execute amended than those focusing connected idiosyncratic aspects of the task.
"The input to the Resnet34 backbone-based web is simply a single-player image," Vats said. "The web outputs 3 probability vectors. The archetypal is the probability of the jersey fig contiguous successful the representation considering each jersey fig successful the dataset arsenic a abstracted class, the 2nd is the probability organisation of the archetypal digit successful the jersey fig and the 3rd is the probability of the 2nd digit successful the jersey number."
The researchers trained their neural network with the weighted sum of the cross-entropy nonaccomplishment of the 3 outputs they focused on. When they tested their network, they recovered that learning some holistic (e.g., '12') and digit-wise (e.g., '1' and '2' successful '12') representations of numbers importantly improved their network's quality to admit jersey numbers. In fact, their multi-task learning attack outperformed different techniques that lone focused connected either the holistic practice oregon digit-wise representations.
"'When the multi-task nonaccomplishment relation web we projected was plugged into a web introduced successful a erstwhile study, it showed a important betterment successful performance," Vats said. "Notably, the multi-task nonaccomplishment relation is besides casual to instrumentality successful a modern heavy learning room (such arsenic Pytorch) and tin beryllium utilized for jersey fig designation successful different sports specified arsenic soccer."
In the future, the neural web developed by this squad of researchers could assistance to automatically place jersey numbers successful sports videos faster and much efficiently. In addition, Vats and his colleagues compiled a caller dataset containing 54,251 annotated images of NHL players and their jersey numbers that could beryllium utilized to bid different techniques for jersey fig and subordinate recognition.
In their adjacent studies, the researchers program to amended their jersey fig and subordinate recognition strategy further. For instance, they would similar to devise a neural web that besides takes into information the determination of crystal hockey players connected the crystal rink erstwhile trying to find their identities.
"The existent survey does not instrumentality temporal discourse into account, truthful our aboriginal enactment volition purpose to amended subordinate recognition by utilizing temporal video information for inferring the jersey fig from broadcast clips," Vats said. "This tin beryllium done done a temporal convolutional web that tin straight enactment connected videos. The projected multi-task nonaccomplishment relation volition beryllium incorporated successful the temporal network."
More information: Multi-task learning for jersey fig designation successful crystal hockey. arXiv:2108.07848 [cs.CV]. arxiv.org/abs/2108.07848
Journal information: arXiv
© 2021 Science X Network
Citation: A multi-task learning web to admit the numbers connected jerseys of sports squad players (2021, September 13) retrieved 13 September 2021 from https://techxplore.com/news/2021-09-multi-task-network-jerseys-sports-team.html
This papers is taxable to copyright. Apart from immoderate just dealing for the intent of backstage survey oregon research, no portion whitethorn beryllium reproduced without the written permission. The contented is provided for accusation purposes only.