Scene Classification for Sports Video Using Transfer Learning

Abstract: The quality of scene classification is particularly crucial because it is a key step in both video summarization and video analysis. This article proposes a strategy to accomplish high-quality scene classification and focuses on several practical implementation gaps over the previous methodologies. We identify five scene types, including batting, bowling, boundary, crowd, and close-up, using cricket as a case study. For scene categorization, we use the ResNet152 model which is pre-trained on imagenet. The proposed method employs new, fully connected layers in an encoder fashion. We also prepared a dataset of cricket video and manually labelled them into five classes which was later used to train the model. The model achieved mean accuracy of 99.15 upon training on the dataset.