A Study on How to Generate Fire Data from Video/Image Using the F-guess and ROI Method

This paper presents the best method for generating fire data and improving the fire recognition rate. At the same time, shorten the labelling time by using fire videos and images when there is a limit to collecting the data and hard to improve the recognition rate with a small label such as a fire. In order to improve the recognition rate and shorten the labelling time, the F-guessed method and the region of interest (ROI) expression method were used to process the data so that the predicted result labelling value and the correct answer value is similar. As a result, data generation increased by about 5.4 times from 5,565 data to 35,633 data compared to the initial labelling task, and mAP@0.5 improved by about 17.6% from 65.9% to 83.5%.