fire-detection-cnn

Why we included this project

This repo targets fire detection in a single frame, with no motion or temporal cues to lean on. It ships several CNN architectures, FireNet and the InceptionV1/V3/V4 variants, each with downloadable pre-trained weights, so you can trade detection accuracy against frame throughput depending on your hardware and how many false alarms you can tolerate. The superpixel-based models go further, splitting each frame into SLIC segments and classifying them individually, which gives you a rough location of the fire rather than just a frame-level flag. The work is backed by two peer-reviewed papers, handy if you need to cite the method or reproduce the reported accuracy in your own evaluation. Teams building safety monitoring or robotics perception systems will find the trade-off guidance between the models a practical place to start.