Classification of Cricket Shots from Cricket Videos Using Self-attention Infused CNN-RNN (SAICNN-RNN)

Arka Dutta, Abhishek Baral, Sayan Kundu, Sayantan Biswas, Kousik Dasgupta, Hasanujaman

Abstract: Millions of people play and enjoy the game of cricket, however, classifying the diverse batting style and postures used frequently by the batsman during a cricket match has always been a difficult proposition. Owing to the immense overlap between postures and various styles of the same shot, it gets extremely harder. Sports experts, trainers, and coaches must learn more about the variety of approaches used by each batsman in both international and local matches in order to guide the team in its entirety and ensure that the training program of cricket players is planned and executed to its fullest potential. The work in this paper thrives on a hybrid deep learning approach that combines convolutional and recurrent neural networks for classifying ten (10) types of cricket shots from match videos. To establish a baseline, a sports CrickShot10 [1] dataset and an open-source cricket video dataset are used. Automatic feature extraction is handled by a hybridized form of convolutional neural network (CNN [11])- recurrent neural network (RNN) combined. Long temporal dependencies are handled by a Gated Recurrent Unit (GRU [12]). It is further improved by adding a Self-attention [20] module that is introduced to the hybrid module to facilitate a semi-supervised approach to extract the key frames from the video. This idea is intended to address the architecture’s inconsistent behaviour while processing somewhat long videos, and their inability to give “correct/relevant” frames priority. When results are compared to other modules, they show good accuracy values. Here we focused on ‘Accuracy’ instead of other evaluation metrics as this is a task of simple classification.

External IDs:dblp:conf/cicba/DuttaBKBDH23