jna-358/fire_actioncam

Fire Actioncam

This dataset is a collection of several real-world fire scenes, introduced by the ECCV paper "Gaussians on Fire: High-Frequency Reconstruction of Flames".

Overview

The dataset consists of 17 real-world scenes of burning paper, cardboard, wood, gasoline, ethanol, and propane. We captured each scene with three regular actioncams, synchronizing them with µs precision using a custom LED pattern.

Property	Value
Scenes	17 (two outdoor locations)
Cameras per scene	3 × GoPro HERO13 Black
Frame rate	400 fps (capture)
Resolution	720 × 1280 (portrait, upright)
Duration per scene	9.5 – 14.9 s (3800 – 5972 frames per camera)
Exposure	1/720 s (692.5 µs), ISO 100 (fixed)
Synchronization	LED pattern + rolling shutter, sub-frame (µs-level) per-frame timestamps
Camera poses	COLMAP-registered, 0.17 – 0.47 px mean reprojection error
Video codec	HEVC (H.265), CQ 18
Total size	≈ 5.1 GB

Getting Started

extract.py (requires numpy + opencv-python) converts scenes into common reconstruction formats:

# all scenes, 300 frames each, every 10th frame (40 fps effective), nerfstudio format
python extract.py --out work --num 300 --step 10 --rectified

# single scene, D-NeRF format, half resolution, camera 1 held out for testing
python extract.py scenes/001 --out work --num 200 --step 20 \
    --format dnerf --downscale 2 --test-cams 1

# Neural-3D-Video style (poses_bounds.npy) / COLMAP model + points
python extract.py scenes/001 --out work_n3dv --num 300 --step 10 --format n3dv
python extract.py scenes/001 --out work_colmap --num 100 --step 40 --format colmap

format	output
nerfstudio	transforms.json (OPENCV model, masks, per-frame time + time_us, train/test lists)
dnerf	transforms_train/test.json (Blender convention, time ∈ [0, 1])
n3dv	poses_bounds.npy (LLFF poses + near/far) + splits.json
colmap	sparse/0/{cameras,images,points3D}.txt + splits.json

parameter	description
scenes (positional)	scene directories to process (default: all under scenes/)
--out	output root; each scene goes to <out>/<scene_id>/ (aborts if it already exists)
--num	number of frames to extract per camera (required)
--start	first frame index (default 0)
--step	take every Nth frame; effective fps = 400/N (default 1)
--cams	camera indices to include (default: all)
--rectified	undistort to a pinhole model
--downscale	integer downscale factor (default 1)
--format	nerfstudio, dnerf, n3dv, or colmap (default nerfstudio)
--test-cams	hold out camera(s) as test set
--test-every	hold out every Nth extracted frame as test
--test-time	hold out a fraction range of the sequence, e.g. 0.8:1.0
--image-format	png (default) or jpg
--jobs	parallel workers (default: number of CPU cores)

Every extraction also writes per-camera per-pixel exposure-offset maps (exposure_offsets.npz, µs) consistent with the chosen geometry (--rectified, --downscale), plus the device masks. Train/test splits are controlled with --test-cams (spatial), --test-every / --test-time (temporal).

Scenes

Scene	Fuel	Description
001	Propane	Low-intensity flame with mild wind from the left
002	Propane	High-intensity flame without wind; detached upper flames
003	Propane	Very low-intensity flame under quiescent conditions
004	Propane	Rapidly growing flame; pronounced vortical structures
005	Propane	Medium-height, wide flame exhibiting rapid growth
006	Propane	Sustained high flame with continuous combustion
007	Wood + Gasoline	Low-intensity flames propagating around wooden logs
008	Gasoline	Medium-scale flames with visible smoke production
009	Cardboard + Ethanol	Small, semi-transparent flames with visible smoke
010	Cardboard + Gasoline	Large flames under wind from the right
011	Cardboard	Small residual flames with visible ash formation
012	Cardboard + Gasoline	Large flames under wind from the right
013	Cardboard	Low-intensity residual flames with visible ashes
014	Paper + Gasoline	Very large flames under wind from the left
015	Paper	Small residual flames with visible ashes
016	Wood + Ethanol	Medium-scale flames; wood mostly unburnt
017	Wood + Ethanol	Small flames; wood partially charred

Data Layout

scenes/
├── 001/
│   ├── 000.mp4          # full sequence video
│   ├── 001.mp4          
│   ├── 002.mp4          
│   ├── 000.mask.png     # synchronization pattern mask
│   ├── 001.mask.png
│   ├── 002.mask.png
│   └── meta.npz         
├── 002/ …
└── 017/
extract.py               # conversion script

We include masks of the synchronization pattern so these can be ignored in reconstruction tasks and the corresponding metrics.

meta.npz Reference

All arrays are indexed by camera c ∈ {0, 1, 2} (matching {000,001,002}.mp4).

key	description
times	per-frame capture time [µs], exposure start of the first sensor row; zeroed at scene start
camera_matrix	intrinsics of the (distorted, upright) videos
dist_coeffs	OpenCV distortion (k1, k2, p1, p2, k3)
R, t	world→camera extrinsics (OpenCV convention, arbitrary scale)
image_size	(width, height) per camera
points3D, points3D_rgb	triangulated static-scene points (world frame) + colors
near, far	robust per-camera scene depth bounds
exposure_us	exposure duration (692.5)
readout_us_per_line	rolling-shutter line delay (2.856)
readout_direction	image axis along which capture time advances (e.g. +x)
sensor_row_of_pixel	expression mapping pixel (x, y) to sensor row (e.g. W-1-x)
rotation_applied_deg_cw	rotation applied to make the sensor image upright
capture_fps, video_fps	400 (capture) vs 60 (container)
serials, capture_datetime	camera provenance
interval_source, led_time_origin_us	trim-interval provenance

Capture & Processing

To obtain µs timing accuracy, we leverage the rolling-shutter effect inherent to the CMOS sensor used by our actioncams. Our custom LED pattern is driven by an ESP32, which controls two segments: (1) a sequence of 16 single LEDs displaying an incrementing gray code and (2) five horizontal LED strips. The gray code gives us the period, while the rolling-shutter effect hands us the phase from the LED strips.

The exposure of pixel (x, y) in frame i of camera c spans

start = times[c, i] + sensor_row(x, y) * readout_us_per_line
end   = start + exposure_us

with sensor_row given per camera by sensor_row_of_pixel (a linear expression in the pixel coordinates; the readout direction is horizontal in the upright videos since the cameras were mounted in portrait orientation). Frame i of the three cameras is aligned to within ± half a frame (≤ 1.25 ms); the exact offsets are in times.

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Citation

@article{nazarenus2025gaussians,
  title   = {Gaussians on Fire: High-Frequency Reconstruction of Flames},
  author  = {Nazarenus, Jakob and Michels, Dominik and Palubicki, Wojtek and
             Kou, Simin and Zhang, Fang-Lue and Pirk, S{\"o}ren and Koch, Reinhard},
  journal = {arXiv preprint arXiv:2511.22459},
  year    = {2025}
}

Acknowledgements

We thank Marvin Voigt, Anton Wagner, and Helge Wrede for their assistance in executing the experiments and Michael Lütten for providing the hardware used for the propane fire setup. Furthermore, the authors acknowledge the financial support of Catalyst: Leaders Julius von Haast Fellowship (23-VUW-019-JVH). Sören Pirk wishes to acknowledge the European Research Council (ERC) who partially funded this research through the ERC Consolidator Grant WildfireTwins (Grant agreement ID: 101170158).