Evaluation of Deep Learning-Based Models for Recognition of Skydiving Formations

This study explores the feasibility of employing deep-learning models to perform approximate live judging for 4-way formation skydiving, a discipline where skydivers perform predefined formations in free fall. Given the time and effort required to judge these formations, our research aims to explore a more instantaneous method for approximate scoring. We created a novel dataset from twelve skydiving competitions and the first six skydiving formations, resulting in 2,946 frames annotated with formation names. We have selected and evaluated the effectiveness of several deep learning architectures, including ResNet-50, EfficientViT, FastViT, and several ConvMixer configurations, in classifying these formations. Our findings reveal that selected pretrained deep-learning models could be used to classify skydiving formations and can extract enough features to achieve high classification accuracy, with the ConvMixer_768_32 model achieving the highest overall F1 score of 0.9865. These results indicate the potential of deep learning applications in automated judging of formation skydiving competitions with potential improvements in spectator experience. Despite the promising outcomes, limitations such as the dataset’s scope and disregard for inference time highlight areas for future research.