HPWREN Fire Ignition images Library for neural network training
HPWREN

This version is fairly old and incomplete!
For now, you can use
https://cdn.hpwren.ucsd.edu/HPWREN-FIgLib/index.html
instead. Check
https://cdn.hpwren.ucsd.edu/HPWREN-FIgLib-Data/index.html
for the latest list of available fire ignition sequences.

The HPWREN Fire Ignition images Library for neural network training

18 August 2020 - last update: 28 July 2024

This page provides access to HPWREN's Fire Ignition image Library (FIgLib), usually based on sequences of wildland fire images as seen from Fixed Field-of-View cameras at HPWREN sites (as per A.C. Clarke: "No machine may contain any moving parts"). It contains more than 400 of such sequences of different fires or different cameras, where a camera may even be separate imagers for color, or monochrome with its higher light and near-IR sensitivity.

This data is provided as-is, no guarantees for anything is included.

The primary purpose of this large archive, many gigabytes in size, is the creation of a baseline data set for fire time-sequences 40 minutes before and after individual fire ignitions for neural network training. The specific focus are AI-based detections of wildland fire ignitions based on content within a single or across multiple images. For illustration purposes full resolution MP4 time-lapse videos per individual data set sequence are included as well. The image file names of the principal data set are constructed as:

origin_timestamp _ offset_(sec)_from_visible_plume_appearance .jpg

for example:

1720453551_-02400.jpg
1720453613_-02338.jpg
. . .
1720455951_+00000.jpg
. . .
1720458232_+02281.jpg
1720458292_+02341.jpg

While we are making the data publicly available, its use requires a credit reference to "http://hpwren.ucsd.edu/" in derivative work. We encourage people and projects working with the data to contact us via the feedback form on the main HPWREN web site about what the archive is being used for. While some of it is just curiosity, knowing about FIgLib uses may help us secure future funding, so projects like FIgLib can continue.

FIgLib was conceived, created and is still maintained by Hans-Werner Braun for the High Performance Wireless Research and Education Network (HPWREN) at UCSD. The objective is to support HPWREN's First Responder collaborators, especially wildland firefighters, as well as neural network researchers who are working towards early fire detection and innovative research towards a better understanding of fire initiations and progression. HPWREN's collaboration with CDF/CalFire began October 2000, and included networked real-time mountain-top fixed-FoV and ptz cameras soon after.

Image data availability

  • Access to the individual sequences, images, and time-lapse videos is provided via the HPWREN-FIgLib-Data pointer. It will show a directory view of the sequences named, e.g., like:

    - 20170708_Whittier_syp-n-mobo-m
    - meaning: YYYYMMDD_firename_camera/imager
    - specifically here the July 8 2017 ignition of the Whittier Fire, as seen from the camera on Santa Ynez Peak near Santa Barbara in the "n"orth direction, using the "m"onochrome chip.

    Clicking one of the sequences will show the contents of the specific directory, including the MP4 time-lapse video and the individual JPEG files utilizing the file name format mentioned above.

    As of July 2024 the whole archive is almost 30 GB in size.

    The Tar directory contains whole image sequences in tgz (tar czf) compressed format.

    Data sets of notable long fire data sequences, along with a large tar file of those sequences are available for comparison purposes, as well as to consider neural networks to interpret fire behavior. The JPEG file names here do not follow the FIgLib naming convention, and just use the image-origin Unix Epoch time stamp.

    To investigate temporal aspects of a fire plume, using motion vectors to estimate movement, the HPWREN-FIgLib-Motionvectors directory contains files along the objectives in the http://hpwren.ucsd.edu/news/20190823/ summary. A large tar file of the motion vector experimental data is available as well.

    Image labels

    For the time being we labeled a few fires with Bounding Boxes, using Jim Davidson's labelem.py python3 script. Some data sets for this are available in CSV and in XML format.

    Regarding labelem.py: It makes things easier to start by pressing "End" to go to the end of a fire data set and work backwards on labeling the images, pulling the bounding box with the left mouse button, and then "Shift-leftarrow" to carry over the bounding box to the previous image in this time series. Don't forget "Save and Exit" at the end to write out the output files. The program has been tested on Ubuntu Linux, and can be run like:

    python3 ./labelem.py "--image=imagedir/*.jpg" --bbfile=bboxfile.csv --screensize 2540x1508 --vocdir=VOC/xmloutfile

    The screensize parameter depends on your screen. While labelem.py generally works, it is not a finished product, so your mileage may vary. Some disclaimers are in the source code.