CricShot10k: A Large-Scale Video Dataset for Cricket Shot Classification

Cricket, one of the most popular sports worldwide, involves a wide variety of batting shots that directly affect match outcomes and play a central role in analyzing player performance. In this context, our study focuses on building an extensive dataset for cricket shot classification from videos and addressing the challenges of existing approaches. We observe that previous works suffer from the time-consuming task of manually trimming match videos into individual shot clips, which has led to relatively small datasets. To overcome this limitation, we propose a novel approach that leverages recurring patterns in cricket broadcasts to automatically split long videos into multiple shot clips. This reframes the task into a form similar to text or image annotation, thereby facilitating the construction of large-scale datasets. Using this method, we build a dataset, called CricShot10k, consisting of 10,086 video clips across 15 shot classes. Our dataset is nearly eight times larger and more diverse than existing ones, covering shots from all formats, and being the first to include women’s and under-19 matches, as well as older matches dating back to 1996. Building on this dataset, we present a modified cricket shot classification framework, CricShotNet, which incorporates two new layers: cropping and segmentation. To identify the best-performing architecture for the subsequent layers, we conduct an extensive comparison across 15 deep feature extractors and 9 variations of temporal feature extractors using GRU, LSTM, and BiLSTM. Our experiments show that EfficientNetV2-S combined with a GRU of 128 units achieve the highest accuracy of 89% on CricShot10k and outperform other state-of-the-art approaches on existing cricket shot video datasets, thus demonstrating the effectiveness of the proposed framework. Our codes and dataset are available at https://github.com/Dihan69/CricShot10k.