ONFIRE Contest 2025: Fire and Smoke Detection in Real-Time on the Edge

Fire detection in surveillance systems represents a critical challenge for protecting human lives, infrastructure and natural ecosystems. Despite advances in deep learning-based approaches, current methods still struggle with the inherent complexity of fire phenomena, leading to detection failures in challenging scenarios and false alarms that undermine system reliability. The ONFIRE Contest 2025 addresses these limitations by fostering international collaboration to develop state-of-the-art real-time fire and smoke detection algorithms optimized for edge deployment on resource-constrained devices. This paper presents the contest framework, dataset and evaluation methodology designed to advance practical fire detection solutions. The contest introduces a scenario-based classification system that organizes detection challenges into four categories based on activity level and viewing distance: Low Activity-Short Range (LA-SR), Low Activity-Long Range (LA-LR), High Activity-Short Range (HA-SR), and High Activity-Long Range (HA-LR). The expanded ONFIRE 2025 dataset comprises over 350 annotated videos from multiple public repositories, representing the largest and most diverse fire detection video collection publicly available. The performance of the competing methods are assessed using the novel Constrained Fire Detection Score (CFDS), which balances detection accuracy with computational efficiency metrics including processing frame rate and memory usage. The results, compared with state of the art approaches evaluated on the same test set, give useful insights for future developments.