WildfireX-SLAM Dataset
- WildfireX-SLAM is a synthetic, large-scale RGB-D dataset designed for simultaneous localization and mapping in realistic forest and wildfire scenarios.
- It leverages Unreal Engine 5 and AirSim for procedural scene generation and automated data capture, simulating dynamic fire, smoke, and varied weather conditions.
- The benchmark evaluates SLAM performance using metrics like ATE, PSNR, and depth error across controlled scenarios, highlighting challenges in dynamic, low-structure environments.
Searching arXiv for the specified paper and closely related work on wildfire SLAM and dynamic Gaussian-splatting SLAM. WildfireX-SLAM is a large-scale, high-fidelity synthetic dataset for simultaneous localization and mapping in large-scale forest and wildfire environments, introduced to address the absence of a comprehensive and high-quality dataset for 3D Gaussian Splatting-based SLAM outside the small-scale indoor scenes emphasized by much recent work (Sun et al., 31 Oct 2025). The pilot dataset contains 5,500+ low-altitude RGB-D aerial images collected over a procedurally generated forest covering 16 km², together with ground-truth camera poses and additional sensing and environmental modalities. It is positioned for wildfire emergency response, forest management, and related UAV and multi-agent robotics tasks, and is accompanied by a benchmark designed to expose the distinctive failure modes of SLAM in forests affected by wildfire dynamics (Sun et al., 31 Oct 2025).
1. Scope and problem setting
WildfireX-SLAM is described as a large-scale, comprehensive, and high-quality synthetic dataset for SLAM in wildfire and forest environments. Its core motivation is methodological: while 3D Gaussian splatting and subsequent variants have led to remarkable progress in SLAM, most recent 3DGS-based systems have concentrated on small-scale indoor scenes. The paper argues that extending 3DGS-based SLAM to large-scale forest scenes has clear relevance to wildfire emergency response and forest management, but that progress is impeded by the lack of suitable data and by the fact that collecting such data in real-world wildfire scenes is costly and technically infeasible (Sun et al., 31 Oct 2025).
The dataset’s reported spatial and operational envelope is unusually specific. It covers 16 km² of procedurally generated forest, contains 5,500+ low-altitude aerial images, and is designed around unmanned aerial vehicles flying below 50 meters. This low-altitude regime is significant because it supports near-ground and aerial SLAM research rather than only high-altitude remote sensing. The intended applications named in the paper include real-time scene understanding, localization, mapping, and navigation in dynamic wildfire conditions; forest inventory, vegetation mapping, and ecological monitoring; and multi-perspective or multi-agent coordination tasks enabled by synchronized aerial and ground views under varied conditions (Sun et al., 31 Oct 2025).
The dataset is also explicitly framed as a pilot dataset. That designation is important in context: the contribution is not only a collection of synthetic scenes, but a controlled infrastructure for studying SLAM under combinations of wildfire, weather, illumination, and terrain conditions that are difficult to standardize in field data.
2. Synthetic environment and acquisition pipeline
WildfireX-SLAM is built with Unreal Engine 5 and leverages the Unreal Engine 5 Electric Dreams Environment Sample Project. The forest environment uses PCG (procedural content generation) and Quixel Megascans assets for realistic trees and terrain. Wildfire phenomena are instantiated with the M5 VFX Niagara Fire System for procedural, high-resolution fire and smoke dynamics. Unreal Blueprints are used for synchronizing vegetation density, terrain slope, lighting, weather, smoke, fire intensity, flame spread, and fog, thereby establishing controllable environmental conditions rather than fixed scene snapshots (Sun et al., 31 Oct 2025).
The acquisition stack modifies and integrates AirSim into Unreal Engine in order to automate realistic sensor data capture from UAVs. The pipeline supports both predefined and manual UAV flight paths, including exploratory, spiral, Z-scan, and target-search trajectories. Cameras and sensors are centrally mounted on the UAV and are adjustable in real time, enabling the capture of detailed trajectories, images, and sensor data. The abstract further notes that the pipeline can collect both aerial and ground views, which broadens the experimental scope beyond a single viewpoint family (Sun et al., 31 Oct 2025).
This design makes controllability a first-class dataset property. Lighting, weather, wildfire type, and wildfire condition are not incidental byproducts of scene generation; they are parameters of the data-generation process. For benchmarking, that matters because sequences can share the same environmental geometry while differing in dynamic or visual conditions, allowing cleaner attribution of performance changes to fire, smoke, fog, or lighting rather than to changes in scene layout.
3. Data modalities and annotation structure
Although the title foregrounds RGB-D, the described release is multi-modal. It provides high-resolution RGB images, dense per-frame depth maps, thermal imagery, normal maps, decomposed reflectance, camera and motion data, and environmental annotations (Sun et al., 31 Oct 2025).
| Modality | Reported content |
|---|---|
| RGB Images | High-resolution images |
| Depth Images | Dense per-frame depth maps (RGB-D) |
| Thermal Imagery | Thermal images |
| Normal Maps | Surface orientation information |
| Decomposed Reflectance | Reflectance decomposition |
| Camera & Motion Data | Ground-truth camera poses , position, orientation, velocity, acceleration |
| Environmental Annotations | Lighting, fog, wildfire state, fire/smoke parameters |
All data streams are sampled at 30 Hz. The pose annotation is given as ground-truth camera poses , and the motion channels include position, orientation, velocity, and acceleration for each frame. The environmental annotations are not merely categorical labels; the paper describes lighting, fog, wildfire state, and fire/smoke parameters as controllable variables, which makes the dataset useful for both supervised evaluation and controlled stress testing of reconstruction, localization, and dynamic-scene filtering (Sun et al., 31 Oct 2025).
For SLAM, this modality mix is consequential. RGB and depth support dense geometric reconstruction; normals and reflectance can support representation learning or photometric decomposition; thermal imagery introduces a sensing channel with direct wildfire relevance; and the motion and environmental channels support analysis of failure under specific visibility or dynamics regimes. This suggests that WildfireX-SLAM is not only a benchmark for camera trajectory estimation, but also a substrate for dynamic-scene understanding and cross-modal perception.
4. Benchmark design and evaluation methodology
The benchmark is organized into two parts. Benchmark A contains two low-altitude UAV sequences, one without wildfire and one with wildfire. Benchmark B defines four challenge scenarios collected under the same environmental geometry but with different dynamic or visual conditions (Sun et al., 31 Oct 2025).
| Scenario | Condition |
|---|---|
| CS1 | Dynamic objects (fire/falling leaves) |
| CS2 | Varying lighting (morning/noon/night) |
| CS3 | Weather effects (no fog vs. fog) |
| CS4 | Combined adverse conditions (wildfire, leaves, fog, poor lighting) |
The evaluated SLAM algorithms are MonoGS and SplaTAM as 3DGS-based methods, Co-SLAM as a NeRF-based baseline, and ORB-SLAM3 as a feature-based baseline. The stated evaluation focus is real-time pose and mapping in highly dynamic, low-structure scenes (Sun et al., 31 Oct 2025).
The benchmark uses trajectory, depth, and photometric rendering metrics. Absolute Trajectory Error is reported as
which measures localization error in meters. Depth error includes Depth L1 and Depth RMSE, with Depth L1 defined as
Photometric rendering quality is evaluated with PSNR, SSIM, MS-SSIM, and LPIPS. The paper explicitly states that lower LPIPS indicates better perceptual similarity, whereas higher PSNR and SSIM indicate better quality reconstruction (Sun et al., 31 Oct 2025).
5. Reported results and characteristic failure modes
Benchmark A shows a marked transition from moderate error in a non-fire forest sequence to severe degradation under wildfire dynamics. In Sequence 1 (No Fire), MonoGS reports 2.5 m ATE, 22.94 PSNR, 0.622 SSIM, 0.455 LPIPS, and 1.07 Depth L1; SplaTAM reports 3.1 m, 22.23, 0.652, 0.428, and 0.83; Co-SLAM reports 6.3 m, 21.54, 0.702, 0.353, and 1.44; and ORB-SLAM3 reports 2.1 m ATE. In Sequence 2 (Fire), MonoGS reports 12.5 m ATE, 13.47 PSNR, 0.562 SSIM, 0.649 LPIPS, and 6.13 Depth L1; SplaTAM reports 13.4 m, 13.34, 0.551, 0.549, and 4.70; Co-SLAM reports 52.8 m, 6.77, 0.412, 0.635, and 7.60; and ORB-SLAM3 reports 9.24 m ATE (Sun et al., 31 Oct 2025).
The paper’s stated observations are that wildfire, specifically dynamic flames and smoke, produces a 2–5× increase in localization error (ATE) and a sharp drop in map quality, with PSNR decreasing by nearly half. It also reports that 3DGS-based methods outperform the NeRF-based baseline Co-SLAM in mapping under dynamics, but still degrade substantially in fire conditions (Sun et al., 31 Oct 2025).
Benchmark B extends this picture to controlled stressors. Under harsh dynamics, fog, or poor lighting, ATE RMSE spikes to m, and Depth RMSE also increases sharply. The paper additionally notes a nontrivial interaction between visibility degradation and rendering metrics: illumination and fog can produce lower photometric error in dark or foggy scenes because there is less detail to render, yet localization still worsens. A common misconception is therefore that lower photometric error necessarily indicates more reliable SLAM; the reported results directly contradict that interpretation in these conditions. This suggests a decoupling between rendering fidelity and pose quality when visibility is severely degraded (Sun et al., 31 Oct 2025).
The failure modes identified in the benchmark analysis are consistent with the reported scene characteristics. Forest and wildfire environments introduce dynamic, non-rigid objects such as fire, smoke, and falling leaves; repetitive and weak texture; variable visibility under morning, noon, night, smoke, and fog; compositional complexity from terrain, vegetation, fire, and occlusions; and multi-agent synchronization demands when overlapping UAV views are used. These are presented as the distinctive challenges of 3DGS-based SLAM in the forest setting (Sun et al., 31 Oct 2025).
6. Significance within wildfire robotics and adjacent research
The paper identifies two principal application domains. For wildfire emergency response, WildfireX-SLAM provides high realism and dynamic variability for safe, pre-deployment testing of SLAM, mapping, and navigation systems in repeatable, risk-free wildfire scenarios; it is also intended to support rapid UAV-based mapping as fire spreads, which is relevant to fireline tracking, personnel safety, and resource deployment. For forest management, the dataset is positioned for change monitoring, vegetation health analysis, and long-term ecological studies. The paper also explicitly highlights swarm and collaborative operations, including UAV and ground-robot scene understanding for wide-area coverage and cooperative firefighting strategies (Sun et al., 31 Oct 2025).
Within the broader 3DGS-SLAM literature, WildfireX-SLAM occupies the role of an evaluation environment rather than an algorithmic tracker or mapper. A relevant adjacent method is WildGS-SLAM, a monocular RGB SLAM system for dynamic environments that predicts a per-pixel uncertainty map from DINOv2 features with a shallow MLP and uses that uncertainty to guide dynamic object removal during both tracking and mapping (Zheng et al., 4 Apr 2025). The challenge scenarios in WildfireX-SLAM, especially dynamic fire, smoke, falling leaves, fog, and variable lighting, suggest a natural benchmark for testing whether uncertainty-aware geometric mapping remains effective in wildfire-specific dynamics rather than in generic urban or indoor motion.
A second adjacent direction is wildfire-aware autonomy. FIRE-VLM introduces a VLM-guided PPO framework for UAV wildfire tracking in a physics-grounded fire digital twin, using dual-view RGB sensing and wildfire-specific semantic alignment scores as reward-shaping signals (Webb et al., 6 Jan 2026). A plausible implication is that WildfireX-SLAM’s emphasis on geometric SLAM under controlled wildfire conditions could be paired with semantic control and exploration frameworks of that kind, especially where active firefront tracking, occlusion by smoke plumes, and large-area aerial coverage must be studied jointly.
Taken together, the dataset, benchmark, and neighboring work delineate a technical agenda rather than a closed problem. WildfireX-SLAM documents that large-scale forest SLAM under wildfire dynamics is constrained not only by dynamic objects and adverse visibility, but also by the mismatch between conventional photometric quality indicators and localization robustness. In that sense, its main contribution is epistemic as much as infrastructural: it provides a controlled basis for measuring where current 3DGS-based, NeRF-based, and feature-based systems fail, and for motivating future methods in wildfire emergency response and forest management (Sun et al., 31 Oct 2025).