FieldSense
Document Type
Abstract
Publication Date
Spring 5-1-2025
Abstract
FieldSense will address the challenge of classifying American football formations, a complex task given the endless variations in team setups. Recognizing that many modern formations are close variants of foundational designs (Wikipedia contributors), the project will compile a dataset focused on the most common American formations (Wikipedia contributors). This approach ensures that even if a formation has not been explicitly labeled, most cases will resemble the foundational formations found within the American Football League. Leveraging the PyTorch deep learning framework, FieldSense will implement a convolutional neural network (CNN) using transfer learning by fine-tuning a pre-trained ResNet-18 model on custom-labeled images (Masood). Custom fully connected layers will be added to capture the subtle nuances of each formation, while data augmentation techniques, such as random cropping, rotations, horizontal flips, and normalization, will enhance the model's robustness (Shu). The training pipeline will utilize a locally stored dataset organized into dedicated directories for training, validation, and testing, with performance monitored via metrics like accuracy, precision, and recall and further refined through early stopping and model checkpointing.
Recommended Citation
Miles, Gregory `25, "FieldSense" (2025). Student Research. 245, Scholarly and Creative Work from DePauw University.
https://scholarship.depauw.edu/studentresearchother/245
Comments
Completed as part of the Computer Science Senior Capstone Project.