Research Webpage about Smoke Detection for Fire Alarm: Datasets


   
Introduction
   

You can down load these videos and images freely only for research purposes. Any commerical use is not allowed before we agree.

Please cite the corresponding article in your publications if the data set helps your research.

   
Smoke detection in video sequences based on dynamic texture using volume local binary patterns
(accepted by KSII Transactions on Internet & Information Systems PDF)
   
Smoke videos used in "Smoke detection in video sequences based on dynamic texture using volume local binary patterns": train_set consisted of 10 smoke videos and 5 non-smoke videos, test_set contained another 10 smoke videos that are similar to the train_set and the same 5 non-smoke videos in the train_set.
   
Deep Domain Adaptation Based Video Smoke Detection using Synthetic Smoke Images
(Fire Safety Journal, 2017, 93:53-59. PDF)(preprint in arxiv PDF)
   
Smoke videos used in "Deep Domain Adaptation Based Video Smoke Detection using Synthetic Smoke Images": The test_set contains 500 smoke images(testsmoke) and 500 non-smoke images(testnon); The train_synthetic_smoke_set are produced by Blender-python.The train_real_smoke_set are captured by ourselves. The background images and non-smoke images are collected from ImageNet. We render each frame of smoke image with a new background image. The parameters of rendering, lighting and wind are set randomly in a certain range for diversity. We give a reference - smoke.blend for parameter sets of smoke rendering. As differnet sets of the parameters influence directly the appearance of synthetic smoke images, these images will be realistic or non-realistic. Experiment showed that the non-realistic synthetic smoke images works just as well as more realistic synthetic smoke images.
   
Domain Adaptation From Synthesis to Reality in Single-model Detector For Video Smoke Detection
(preprint in arxiv PDF
   
The source (synthetic smoke samples) dataset: synthetic_set contains 2549 images.The target (real smoke samples) dataset: real_set contains 30000 images.