Mirror-fire-dataset

We take a heritage building within China as the research object. It happens to be a provincial-level cultural relic protection unit that is troubled by fire prevention monitoring. Due to the particularity of heritage buildings, it is impossible to conduct actual fire tests inside, and relevant cultural relic protection laws and regulations do not allow such actions either. Therefore, we use 3DMAX to construct virtual fire scenes for it for subsequent fire detection experiments. By setting and adjusting wide-angle mirrors in the model and then using the rendering method of 3DMAX, we can obtain the corresponding scene images. In this experiment, we have set up four target monitoring areas and one non-interest area. Through adjustment, the indirect vision in the mirror has been aligned with the target monitoring area. Meanwhile, due to the position of the ceiling of the venue, there is no through line. The ceiling is made of non-combustible concrete material and is also far away from the crowd, so the probability of a fire occurring first in this area is nearly zero. Therefore, we did not set up a wide-angle mirror to observe this area, that is, this area is not included in the indirect vision and is used as a non-interest area. In this non-interest area, in a small amount of data, we will hang some flags as some interference factors in the detection to test the anti-interference ability of the detection model. This is also a real situation in the actual operation of the venue. Then, we pointed the camera at the main entrance of the building to collect images, thus forming this fire data set based on indirect vision. This data set contains a total of 800 fire pictures, saved in JPG format with a pixel size of 1645×2493. Meanwhile, in order to verify the ability of the model to resist noise interference in the non-interest area, 138 of them have flags added as noise in the non-interest area.