WildfireX-SLAM v2.0

Synthetic low-altitude UAV sequences over a forest with active fires — RGB, depth, normals, materials and exact poses — for SLAM, 3D Gaussian Splatting and robust perception under smoke, flames, falling leaves, fog and changing illumination.

v2.0 follows the benchmark structure of v1.0 (MMM 2026 paper; differences below) and was collected with the open wildfirex toolkit on simulator WildfireX-Sim 1.1: six fire sites, canopy-following flights just above the trees, float depth, exact camera poses, 30 Hz with 0.1 m per frame, runtime-controlled conditions and automatic quality checks. Every sequence can be regenerated with wildfirex collect wildfirex_slam_v2.yaml --only <id> (toolkit: pip install wildfirex; code, config and docs in the WildfireX-SLAM repository; project page: zhicongsun.github.io/wildfirexslam).

Sequences	19 (sequences/<site>-<number>, sites ed01 … ed06)
Frames	16,572 RGB-D frames at 640×480, 30 Hz
Modalities	RGB, depth (cm), normal, albedo, roughness, specular, AirSim object-ID "infrared"
Ground truth	camera-to-world poses (TUM + Replica formats), intrinsics, velocity/acceleration, fire-kit positions
Flight altitude	follows the surveyed canopy with a minimum clearance of 12 m (spirals of A) or 15 m (all others); per-sequence median 29–42 m above ground, 17–24 m above the canopy
Simulator	WildfireX-Sim 1.1 (Electric Dreams, UE 5.2, AirSim-WildfireX plugin; + movable fire kits, moon)
Collected with	wildfirex 0.1.0 (wildfirex_slam_v2.yaml)
Size	≈ 62 GB (PNG in one zip per modality per sequence; RGB + depth ≈ 17.5 GB)

Benchmarks (splits/wildfirex_slam_v2_0.json)

Benchmark	Sequences	Varied factor	Trajectory
A	ed01-0001 … 0004, ed02-0001 … 0004	no fire vs. fire, two sites	T01 spiral around the fire (42 m away), T02 lawn-mower survey 55 m beside the fire looking at it
CS1	ed03-0001 … 0004	falling leaves	T03 search flight towards the fire (ends 45 m from it), T04 spiral
CS2	ed04-0001 … 0004	morning / noon / dusk / night	T04 spiral
CS3	ed05-0001, 0002	fog 0 vs 0.7	T02 side survey
CS4	ed06-0001	fire + leaves + fog 0.35 + night	T04 spiral

Paired sequences share exactly the same trajectory and fire. Per-sequence conditions are in sequences.csv and each meta.json.

Fire: three fire kits (ground flames + smoke) form a ≈ 24 m fire front at every site; at ed01 the level's original fire and smoke column are kept as well. At dusk a low sun lights the canopy from the opposite side to morning and the fire glows faintly; at night a weak moon and the fire light the scene, so it is dark away from the fire.

Conditions

Download

Each sequence folder holds its poses and metadata as small text / json files plus one zip per modality (rgb.zip, depth.zip, normal.zip, albedo.zip, roughness.zip, specular.zip, ir.zip), so you can fetch only what you need. RGB + depth for all sequences is ≈ 17.5 GB of the ≈ 62 GB total.

pip install -U huggingface_hub
# RGB-D, poses and metadata of all 19 sequences (~17.5 GB): skip the other modalities
hf download zhicongsun/WildfireX-SLAM --repo-type dataset --local-dir WildfireX-SLAM   --exclude "*/normal.zip" --exclude "*/albedo.zip" --exclude "*/roughness.zip" --exclude "*/specular.zip" --exclude "*/ir.zip"
# one complete sequence
hf download zhicongsun/WildfireX-SLAM --repo-type dataset --local-dir WildfireX-SLAM --include "sequences/ed01-0002/*"
# everything (~62 GB)
hf download zhicongsun/WildfireX-SLAM --repo-type dataset --local-dir WildfireX-SLAM

Unzip in place (each <modality>.zip extracts to <modality>/000000.png ... inside its sequence folder):

python -c "import glob,os,zipfile; [zipfile.ZipFile(z).extractall(os.path.dirname(z)) for z in glob.glob('WildfireX-SLAM/sequences/*/*.zip')]"

Format

Conventions — read before using TUM / Replica loaders

Depth	16-bit PNG, planar depth in centimetres: metres = value / 100 (depth_scale: 100 in intrinsics.yaml); 0 = invalid. This is not TUM's 5000 — set the depth scale of your loader to 100, otherwise all depths come out 50× too small. Centimetres are needed because depth reaches several hundred metres outdoors (TUM's 5000 would cap it at 13 m).
Camera frame	OpenCV: x right, y down, z forward — the same as TUM and Replica.
Poses	camera-to-world. groundtruth.txt: TUM format timestamp tx ty tz qx qy qz qw; traj.txt: Replica format, one row-major 4×4 matrix per line.
World frame	AirSim NED: x north, y east, z down (metres, right-handed). All 19 sequences share one world frame. Trajectory metrics (ATE/RPE with alignment) are unaffected; viewers that assume z up show the scene upside down.
Intrinsics	pinhole, no distortion: fx = fy = 320, cx = 320, cy = 240 (640 × 480).

To get a z-up world (x north, y west, z up), left-multiply every pose; the camera frame stays OpenCV:

import numpy as np
F = np.diag([1.0, -1.0, -1.0, 1.0])   # NED -> NWU (z up), still right-handed
c2w_zup = F @ c2w

Full file list: docs/data_format.md in the WildfireX-SLAM repository. In short: depth/*.png are uint16 centimetres (0 = invalid); poses are OpenCV camera-to-world with the world in AirSim NED (metres); intrinsics.yaml gives fx = fy = 320, cx = 320, cy = 240. meta.json records ned_origin_ue_cm (the AirSim NED origin in Unreal world coordinates), the fire-kit positions and all versions, so sequences from different sites can be put in one world frame.

Sequence ids

<site>-<number>, e.g. ed04-0003: ed = the Electric Dreams map, 01–06 = fire site (ed01 is the level's original fire, ed02–ed06 were chosen in the forest from the canopy survey), 0001… = running number at that site. Conditions (fire, light, fog, leaves, trajectory) are fields in sequences.csv and meta.json, never part of the id, so new conditions never rename data.

Modalities (all 640×480, same frame)

folder	content
rgb	final rendered colour image
depth	planar depth (distance along the optical axis), uint16 PNG in cm, 0 = invalid / > 655 m; includes smoke
normal	surface normal from the G-buffer, encoded as RGB
albedo	base colour of the surface (material colour without lighting, shadows or fog) from the G-buffer
roughness	material roughness (G-buffer), grey
specular	material specular (G-buffer), grey
ir	AirSim object-ID "infrared" image (grey value per object), not a physical thermal image

Differences from v1.0 (MMM 2026 paper) — please read

• Scale and views: v1.0 has 5.5k RGB-D images over a 16 km² forest map, with aerial and near-ground views. v2.0 has 16,572 frames at six fire sites within ≈ 600 m of the central forest, aerial views only.
• Altitude: v1.0 flies below 50 m above ground, Benchmark A within 15 m above ground. v2.0 follows the canopy at least 12–15 m above the trees: 26–53 m above ground, per-sequence median 29–42 m.
• Benchmarks: A has 8 sequences at two sites (v1.0: 2). CS1 varies falling leaves with the fire on, at one site (v1.0: with / without fire in two environments, with / without leaves). CS2 adds dusk to morning, noon and night. CS3 (fog) and CS4 (fire + leaves + fog + night) follow v1.0.
• Modalities: v1.0 includes thermal images. v2.0 has RGB, depth, normal, albedo, roughness, specular and an object-ID image (ir, not thermal); it has no thermal images.
• Fire: ed02–ed06 use fire kits added to the level (WildfireX-Sim 1.1); ed01 uses the level's original fire.
• Views of the fire: the camera looks at the fire from 42–110 m (spiral, side survey, search); a pass right over a fire puts the camera inside the smoke column (> 50 % of the image), so v2.0 has no nadir pass over the flames.
• Depth content: depth includes smoke, flames and falling leaves (same instant as RGB); foliage moves with wind.

Quality

All 19 sequences pass the automatic checks (qa.json, sequences.csv): no missing frames, valid depth 96.7–100 %, no camera-inside-geometry frames (no frame with > 20 % of pixels closer than 0.5 m), median relative depth reprojection error between consecutive frames 0.08–0.42 % (higher where smoke, which is part of the depth, moves), and a consistent first frame (5 pre-roll captures are dropped before frame 0; the share of inconsistent pixels between frames 0 and 1 matches that of later frame pairs).

Note: fog 0.7 (CS3) is dense; at night (CS4) the fire light makes fog glow, so CS4 uses fog 0.35.

License

WildfireX-SLAM Dataset License: the terms of CC BY 4.0 (any use, including commercial, with attribution) plus one restriction — no use in, for the development of, or as training input to generative AI programs. The restriction passes on the "NoAI" terms of the third-party Unreal Engine assets depicted in the data (Electric Dreams Environment by Epic Games, with its Quixel Megascans assets; M5 VFX Vol2 Fire and Flames). SLAM, pose and depth estimation, 3D reconstruction and rendering of the captured scenes, detection, segmentation and benchmarking are not affected. No rights to the underlying assets are granted.

Unreal® is a trademark or registered trademark of Epic Games, Inc. in the United States of America and elsewhere.

Citation

@inproceedings{sun2026wildfirex,
  title={WildfireX-SLAM: A Large-Scale Low-Altitude RGB-D Dataset for Wildfire SLAM and Beyond},
  author={Sun, Zhicong and Lo, Jacqueline and Hu, Jinxing},
  booktitle={International Conference on Multimedia Modeling},
  pages={533--546},
  year={2026},
  organization={Springer}
}