The dataset
The Large Fire Dataset improved with Negative samples (LFDN) is a dataset of images specifically designed for training effective frame-wise flame and smoke detectors. The images present variations in terms of illumination conditions, acquisition range, presence of moving objects, and application scenario (e.g., forest fires or urban ones). Furthermore, it has been collected with the idea to include, together with representative positive samples, negative samples with fire-like objects, including various environmental situations (sunlight, clouds, fog), illumination conditions (reflections, glares) and white or red objects (headlights, dust, clothes, flags). The dataset consists of 36,554 images, annotated with a bounding box surrounding each flame or smoke and the corresponding category (flame/smoke). In total, there are 8,070 images with flames, 15,330 images with smoke, 9,000 images with both and 4,154 negative samples.
References
If you use this dataset please cite:
- Fire detection in videos combining a deep neural network with a model-based motion analysis @ARTICLE{gragnaniello2024_fire,
title = {Fire detection in videos combining a deep neural network with a model-based motion analysis},
author = {Diego Gragnaniello, Antonio Greco, Carlo Sansone, Bruno Vento},
year = {2024},
journal = {Submitted to Neural Computing and Applications},
}
Download
In order to download the whole dataset click here.
Help
If you have any problems, do not hesitate to contact us here.