Year: 2025
Type: preprint
Abstract: Fire detection in industrial and warehouse settings is a high-stakes task, yet collecting large-scale datasets of real fire imagery is costly, dangerous, and often infeasible. This preliminary report (completed December 22, 2023) describes the development of a Unity-based synthetic dataset for fire detection, the design of small-scale classification and detection experiments, and the public release of synthetic images and training scripts. The goal was to explore the feasibility of synthetic-to-real transfer for early-stage detection under varied lighting and camera conditions. Quantitatively, the TinyVGG classifier achieved 96% test accuracy on synthetic data, and the YOLOv5 detector reached [email protected] = 0.88 on the synthetic validation set; in a live demo the detector correctly identified 10/10 real-world fire images. This work was first released publicly on GitHub on December 18, 2023 and subsequently presented as a poster in January 2024.
DOI: 10.31224/5503
Author: Christopher O'Hara
Institution: The University of Tokyo
Language: English
Subfield: Safety, Risk, Reliability and Quality
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4414790623
Cites:
Cited by:
Related to: