FireUnitData: Building Unit Fire Detection
- FireUnitData is an annotated UAV dataset that focuses on fire localization at individual building units rather than broad scene-level detection.
- It combines real drone imagery with synthetic enhancements like motion blur, brightness variation, and generated flames to simulate varied fire scenarios.
- The dataset’s diverse facade categories and targeted annotations aim to improve resource allocation and emergency response through precise detection.
FireUnitData is a drone-view building-unit fire detection dataset introduced to address the lack of annotated data at the granularity of individual building units in fire scenes. Its stated purpose is unit-level localization in building fire scenes rather than broad scene-level fire detection, with an emphasis on identifying which exact unit on a building facade is on fire so that responders can route personnel more efficiently, allocate resources better, and issue more targeted evacuation guidance (Zhai et al., 5 Aug 2025). In the same literature landscape, the term also supports a broader interpretation as structured fire-response data centered on operational units, incidents, or spatial cells, but the canonical named resource is the building-unit dataset released at https://github.com/boilermakerr/FireUnitData (Zhai et al., 5 Aug 2025).
1. Definition and problem setting
FireUnitData is presented as an annotated dataset of building units captured by drones and enhanced through multiple synthetic transformations. The paper frames the problem as a gap between existing fire datasets, which mainly support coarse fire recognition, and the operational need for building-unit-level localization. Existing datasets cited there support tasks such as fire/smoke object detection, aerial burn detection, multispectral aerial fire observation, or remote-sensing fire detection, but they do not provide annotations at the granularity of apartments, windows, or facade units (Zhai et al., 5 Aug 2025).
The dataset therefore targets a specific computer-vision problem: building-unit fire localization from UAV imagery. Its core design combines real building facade imagery with synthetic fire insertion and UAV-condition augmentation. The motivation is explicitly twofold: the lack of annotated unit-level fire data and the difficulty and risk of collecting real fire data at scale. The authors state that the dataset “can effectively improve the generalization ability of fire unit detection,” especially under multi-scenario conditions, although the provided manuscript excerpt does not include a quantitative benchmarking section that substantiates this claim numerically (Zhai et al., 5 Aug 2025).
A broader reading of the term is also visible in adjacent work. In FireCommander, “FireUnitData” is effectively the structured state associated with firefighting units, fire evolution, facilities, user commands, and score signals over time, although the term is not the title of that work (Seraj et al., 2020). This suggests that FireUnitData can denote either a specific UAV facade dataset or, in a wider research vocabulary, unit-centered fire data representations.
2. Dataset composition and scene taxonomy
The concrete dataset statistics are narrowly specified. FireUnitData contains 1,978 images, split into 1,777 training samples and 201 test samples, and each sample comprises an original image and an annotation file in TXT format (Zhai et al., 5 Aug 2025).
| Property | Value |
|---|---|
| Total images | 1,978 |
| Training set | 1,777 |
| Test set | 201 |
The paper does not explicitly mention a separate validation split, does not explicitly state the number of classes, does not define the exact TXT annotation schema, and does not explicitly state image resolution. It also does not explicitly specify whether annotations mark only flaming building units, all building units, facade regions, or fire regions. The text strongly implies building-unit fire localization, but finer annotation semantics remain underspecified (Zhai et al., 5 Aug 2025).
The dataset’s visual diversity is structured around six regions and eight building types or building facades. These are grouped into four broad facade categories: high-rise buildings, multi-story buildings, row houses, and standalone buildings. The paper further characterizes these categories in architectural terms. High-rise buildings are usually 10+ stories and contain many vertically stacked units with dense and highly similar structure. Multi-story buildings have fewer than 10 stories, fewer vertical units, and more variability in window structures and unit arrangement. Row houses consist of multiple adjacent horizontally connected units with a wide facade and relatively regular arrangement. Standalone buildings are independent, with narrower facades and mostly consistent unit forms with occasional size differences (Zhai et al., 5 Aug 2025).
This categorical design implies that the dataset was assembled to span different facade layouts, unit densities, and structural regularities. A plausible implication is that the dataset is intended to challenge detectors not only with fire appearance, but also with the geometric regularities and irregularities of facade-unit decomposition.
3. Construction pipeline and synthetic enhancement
FireUnitData is explicitly a hybrid real-plus-synthetic dataset. The starting point is real UAV-captured building imagery obtained with DJI drones. The authors collected high-definition images of eight different building facades from six regions, extracted facade regions, and removed interference from other buildings and the background so that the retained imagery emphasizes the target building-unit regions (Zhai et al., 5 Aug 2025).
The construction workflow then applies three realism-oriented enhancement stages. First, the images are modified to simulate drone-view imaging conditions. Motion blur is synthesized through a convolution-based algorithm implemented with OpenCV’s filter2D, with parameter-controlled blur kernel size and weight distribution. The paper explains this as suppressing high-frequency components while preserving low-frequency structural information, with the intended effect of matching the frequency-domain attenuation characteristics of real motion blur under hovering, cruising, and maneuvering conditions (Zhai et al., 5 Aug 2025).
Second, brightness variation is introduced in HSV color space. The image is converted from BGR to HSV, linear scaling is performed on the V channel, and the result is converted back from HSV to BGR. The stated purpose is controllable brightness adjustment and continuous illumination variation while preserving hue and saturation more effectively than raw RGB perturbation (Zhai et al., 5 Aug 2025).
Third, synthetic fire is inserted using a large image generation model. The workflow starts from an original building image, randomly generates irregular mask regions, treats these masks as target fire locations, and uses a large-scale image generation model to generate flames inside those masked regions. The generation is intended to include flame characteristic simulation and natural integration with the building structures. The paper does not specify which large model was used, whether it was inpainting, diffusion, GAN-based, or prompt-driven, nor whether masks were constrained to window or unit locations (Zhai et al., 5 Aug 2025).
Taken together, the pipeline is: collect real UAV imagery, extract facade regions, preserve scene diversity across the four building categories, apply illumination augmentation, apply motion blur augmentation, generate synthetic fire effects with random irregular masks, and produce annotated samples with image plus TXT label file (Zhai et al., 5 Aug 2025). This suggests a dataset design centered on controllable realism rather than large-scale real-fire acquisition.
4. Annotation semantics and data representation
Annotation information is among the least fully documented parts of the resource. What is explicit is limited: each sample includes an annotation file in TXT format; the dataset is an annotated dataset of building units captured by drones; and the intended task is building-unit fire detection or unit-level localization (Zhai et al., 5 Aug 2025).
By contrast, several aspects are not defined in the provided text: the class list, the exact TXT syntax, whether the labels follow YOLO convention, whether annotations are bounding boxes, polygons, points, or masks, whether all units are labeled or only fire-positive ones, whether fire and unit are separate categories, and whether negative samples without fire are present (Zhai et al., 5 Aug 2025). The most plausible interpretation is that the annotations are detector-oriented bounding boxes for target units, but this remains an inference rather than a paper-stated fact.
This under-specification is notable because other contemporary fire datasets document their annotation targets more concretely. WIT-UAS, for example, stores hand-labeled LWIR images with bounding boxes for humans and vehicles in Pascal VOC-style minX, minY, maxX, maxY, class format (Jong et al., 2023). DetectiumFire explicitly states bounding box annotations for fire regions and supplements them with fire-focused textual prompts and a four-level severity taxonomy (Liu et al., 4 Nov 2025). Relative to those resources, FireUnitData is specialized by scene and task but lighter in schema specification (Zhai et al., 5 Aug 2025).
A plausible implication is that the manuscript prioritizes dataset motivation and generation methodology over benchmark formalization. For reproduction and comparative evaluation, this makes repository inspection operationally important.
5. Relation to adjacent fire-data paradigms
FireUnitData belongs to a wider family of resources that organize fire-related information around operational or analytically meaningful units, but it occupies a distinct niche. Its unit is the building facade unit viewed from a drone, and its target is unit-level fire localization in built environments (Zhai et al., 5 Aug 2025).
Other works define different unit granularities. WIT-UAS centers on crew assets—humans and vehicles—from aerial thermal imagery in prescribed wildland fire environments, with ROS metadata and LWIR images as the primary unit-level records (Jong et al., 2023). PyroGuardian treats the firefighter as the core telemetry unit and fuses per-firefighter identity, GPS location, direction, ambient temperature, heart rate, pulse, SpO, and body temperature into a unit dashboard for incident command (Kaplan et al., 2024). FireCommander formalizes unit-level data as agent telemetry, wildfire states, facility states, and command logs in a probabilistic multi-agent wildfire-response environment (Seraj et al., 2020).
At larger spatial scales, the “unit” becomes an incident or a grid cell. WildfireIA defines one retained FPA-FOD wildfire record as the event unit and attaches discovery-time environmental and contextual features for initial attack failure prediction (Xu et al., 14 Jun 2026). IberFire uses the 1 km 1 km 1 day cell-day tuple as its atomic observation for wildfire occurrence modeling in Spain (Ercibengoa et al., 1 May 2025).
This comparison clarifies the specificity of FireUnitData. It is neither an operational suppression log, nor a firefighter telemetry stream, nor an incident table, nor a spatio-temporal risk datacube. It is a purpose-built UAV vision dataset for building-unit fire localization. This also distinguishes it from broader multimodal fire benchmarks such as DetectiumFire, which spans 22.5k high-resolution fire-related images and 2.5k real-world fire-related videos with both bounding boxes and textual prompts across indoor, outdoor, controlled, and uncontrolled fires (Liu et al., 4 Nov 2025).
6. Limitations, biases, and research significance
Several limitations are evident from the paper’s own reporting. First, the dataset is modest in size at 1,978 images. Second, it relies substantially on synthetic components: synthetic fire insertion, simulated blur, and simulated brightness changes. This creates a synthetic-to-real domain gap, since real emergency scenes may involve smoke, reflections, occlusion, structural damage, and fire-light interactions not fully captured by mask-based flame insertion (Zhai et al., 5 Aug 2025).
Third, scene diversity is broader than a single building type but still limited to six regions and eight building facades. The paper itself suggests possible generalization challenges to very different facade materials, nighttime scenes, dense smoke occlusion, weather conditions, oblique viewpoints, thermal imagery, or severe fire spread across multiple units (Zhai et al., 5 Aug 2025). Fourth, annotation semantics remain ambiguous. Because the manuscript does not define the exact label syntax or taxonomy, benchmark reproducibility and cross-paper comparability are constrained.
A further limitation is the absence of reported detector architectures, training settings, evaluation metrics, baseline comparisons, or ablation studies in the provided text. Unlike DetectiumFire, which reports object-detection results such as YOLOv11m mAP, cross-dataset transfer, and multimodal reasoning improvements (Liu et al., 4 Nov 2025), or WIT-UAS, which reports mAP for SSD300 and YOLOv3 under urban-to-wildfire transfer settings (Jong et al., 2023), FireUnitData’s manuscript excerpt does not provide quantitative experimental characterization (Zhai et al., 5 Aug 2025).
Its research significance therefore lies primarily in task definition and data curation. FireUnitData fills a narrowly defined but operationally consequential gap: identifying the exact building unit involved in a fire from drone imagery. In the broader fire-data ecosystem, this makes it a specialized complement rather than a general benchmark. A plausible implication is that its greatest value is as a task-specific supplement for UAV-based building inspection and rescue support, especially when combined with additional real fire imagery, domain adaptation, or evaluation on operational drone footage (Zhai et al., 5 Aug 2025).
