Network architecture for the player identification model. The networks accepts a player tracklet as input. Each tracklet image is passed through a ResNet18 to obtain time ordered features F . The features F are input into three 1D convolutional blocks, each consisting of a 1D convolutional layer, batch normalization, and ReLU activation. In this figure, k and s are the kernel size and stride of convolution operation. The activations obtained from the convolutions blocks are mean-pooled and passed through a fully connected layer and a softmax layer to output the probability distribution of jersey number p jn .

Context in source publication

Context 1

... to the bounding box o i is input into a backbone 2D CNN, which outputs a set of time-ordered features {F = {f 1 , f 2 .....f n }f i ∈ R 512 }. The features F are input into a 1D temporal convolutional network that outputs probability p ∈ R 86 of the tracklet belonging to a particular jersey number class. The architecture of the 1D CNN is shown in Fig. ...