Comparison of two semantic segmentation databases for smoke detection

Researchers have found strong correlation between warm summer and the frequency and intensity of fires around the world. Climate models due to global warming tells us that average summer temperature will increase drastically in the next few decades entailing an increase of wildfire. Computer vision is a good tools to detect and locate an incipient fire and prevent a rapid spread of fire destroying huge forest areas as in Australia or Brazil. Smoke is the first clue of an incipient fire that can be detected by a camera to warn firemen to act as quickly as possible. Convolutional neural networks and semantic segmentation can achieve this task by giving location and scale of the fire to firemen. In order to efficiently train this type of network architectures, we need a database composed of many images and corresponding masks. The complexity of the smoke in terms of shape, texture, color and intensity is difficult to segment properly. The diversity of smoke types in the image database is crucial for generalizing prediction in real-world circumstances. Numerous research papers proposed new network architectures for segmenting smoke in visible images spectrum and tested the accuracy of the segmentation on their database. Database that, for the most of the time, was not available. This article deals with comparison of a network performances on two smoke databases and highlight the importance of a rich images database in terms of quality rather than quantity.