CAS-FD AI spots real fouls vs dives with 86% accuracy

A team of researchers, led by Md. Jahidul Islam, has released CAS-FD (Contact-Aware Temporal Sampling for Foul vs Dive recognition), a new computer-vision pipeline that tackles one of football's toughest officiating challenges: telling a genuine foul from a theatrical dive using only a single broadcast camera angle. The system, published on arXiv, introduces a contact-aware sampling strategy that concentrates the model's attention on the exact moment of physical contact rather than treating all frames equally. This temporal focus allows the model to recognize subtle contact-specific cues that standard frame-by-frame methods miss.

CAS-FD builds on a new balanced dataset of 600 broadcast clips annotated for fouls and dives, and it achieves 86.0% accuracy with a macro-F1 of 0.860 on the held-out test split—a 12 percentage-point improvement over contact-unaware alternatives, with an even larger gap on unseen data. The researchers also performed a component-wise evaluation against human annotations to identify exactly where and why the system succeeds and fails. By releasing both the dataset and code, they provide a grounded, reproducible benchmark for fine-grained contact-event recognition in broadcast football, opening the door for practical VAR-assist tools and automated officiating aids.