D-Fire: A Fire and Smoke Dataset

• D-Fire Dataset is a large-scale, multi-class image collection designed for fire and smoke detection research and practical fire management applications.
• It comprises 21,527 RGB images with 26,557 annotated bounding boxes across four categories, offering robust benchmarks for object localization and classification.
• The dataset’s diversity in imaging conditions and annotations supports the development of advanced computer vision models with improved accuracy and lower false alarms.

The D-Fire Dataset is a large-scale, multi-purpose image dataset curated for the advancement of automated fire and smoke detection using computer vision and machine learning methodologies. It is distinguished by its balanced and challenging image composition, detailed bounding box annotations, and real-world capture conditions, making it a critical resource for benchmarking and developing fire management and safety technologies.

1. Dataset Composition and Annotations

The D-Fire Dataset encompasses 21,527 RGB images with a standardized spatial resolution of 416×416 pixels. Images are organized into four distinct classes, each reflecting the presence or absence of fire and smoke phenomena:

Category	Image Count	Annotations (Bounding Boxes)
Fire	1,164	14,692 (“fire”)
Smoke	5,867	11,865 (“smoke”)
Fire and Smoke	4,658	Both “fire” and “smoke”
None	9,838	0

Each annotated bounding box is labeled as “fire” or “smoke,” totaling 26,557 object instances. The dataset’s structure supports multi-class and multi-instance detection paradigms, providing a robust basis for both object localization and classification tasks.

2. Imaging Modalities, Resolution, and Provenance

All images in D-Fire are in the RGB spectrum, ensuring compatibility with standard deep learning pipelines. Their uniform 416×416 resolution enables consistent training and evaluation for detection algorithms and reduces the need for preprocessing such as resizing or cropping.

Image sources are heterogeneous, including:

• Legal fire simulations at the Technological Park of Belo Horizonte, Brazil
• Surveillance cameras from state parks and university campuses (e.g., Universidade Federal de Minas Gerais, Serra Verde State Park)
• Web-sourced imagery to incorporate diverse scenarios and visual ambiguities

This diversity promotes significant intra-class variability, encompassing a broad spectrum of weather, visibility, and background conditions, as well as practical challenges such as occlusions (e.g., insects on the lens) or adverse lighting.

3. Experimental Use and Benchmarking

The D-Fire Dataset is purpose-built for object detection tasks in applied fire management. It has been used as a standard benchmark for evaluating models such as YOLOv8, DeepLab-V3, ResNet-50, and compact architectures like YOLOv8n and YOLOv5n (Boroujeni et al., 17 Mar 2025, Joshi et al., 11 Oct 2025).

Key performance metrics, computed in accordance with accepted object detection protocols, include:

• Accuracy:

$$\operatorname{Accuracy} = \frac{\mathrm{TP} + \mathrm{TN}}{\mathrm{TP} + \mathrm{TN} + \mathrm{FP} + \mathrm{FN}}$$

• Intersection over Union (IoU):

$\operatorname{IoU} = \frac{|A \cap B|}{|A \cup B|}$

where $A$ is the predicted bounding box and $B$ is the ground truth.

• Mean Average Precision (mAP):

$$\operatorname{mAP} = \frac{1}{N} \sum_{i=1}^{N} \text{AP}_i$$

Reported both at a fixed IoU threshold (e.g., [email protected]) and across a range (mAP@[.50:.95]) to assess overall and stringent detection performance.

Experimental results have demonstrated that D-Fire presents sufficient challenge to discriminate between models’ capacity to distinguish not only fire from non-fire, but to disentangle fire and smoke in complex, ambiguous scenarios.

4. Distinctive Features and Research Contributions

D-Fire’s design confers several unique benefits:

• Multi-Class Labeling: By distinguishing “fire,” “smoke,” “fire and smoke,” and “none” in both image and object-level annotations, D-Fire enables nuanced modeling that improves over traditional binary detection datasets (Boroujeni et al., 17 Mar 2025).
• Environmental and Visual Diversity: Inclusion of natural, urban, and simulated scenes accommodates training under diverse deployment conditions, reducing domain gap in real-world deployments.
• Challenging Negatives: The inclusion of “none” images with visual features (objects, colors, textures) susceptible to false positives—such as lamp lights or sun glare—fosters development of robust classifiers with lower false alarm rates.
• Open Licensing: The dataset is released under a Creative Commons Zero (CC0) Universal License, supporting broad dissemination and unrestricted reuse.

5. Real-World Applications and Technology Development

The dataset is chiefly intended to support the training, evaluation, and deployment of object detection models vital for fire management systems, including:

• Real-time fire and smoke detection for early warning and surveillance
• Algorithm robustness evaluations in scenarios with high intra-class variation
• Benchmarking the impact of post-detection uncertainty-aware frameworks in compact models deployed on resource-constrained devices (e.g., UAVs, IoT, CCTV) (Joshi et al., 11 Oct 2025)

A salient example is the development and evaluation of an uncertainty-aware post-detection framework that refines detection confidences by integrating model uncertainty (estimated via dropout-based variance) and domain visual features (HSV color histograms, edge, and texture descriptors) (Joshi et al., 11 Oct 2025). On D-Fire, this approach increased YOLOv8n’s mAP@50 from 0.625 to 0.651 and precision from 0.712 to 0.845, with only a modest increase in per-image processing time (12.78 ms to 20.15 ms).

6. Limitations and Challenges

While D-Fire’s breadth and annotation detail afford comprehensive benchmarking, certain challenges are noted:

• RGB-only Modalities: Absence of thermal infrared and other non-RGB channels precludes some multi-modal detection techniques.
• Bounding-Box Level Annotations: Pixel-level segmentation is unsupported, limiting its direct applicability to semantic segmentation algorithms.
• Scene Variability: Although the dataset is diverse, specific environmental conditions or extreme fire/smoke events may remain underrepresented.

A plausible implication is that extending D-Fire with additional sensor data or annotation types could further enable research into multi-modal and cross-domain generalization.

7. Impact and Open Research Directions

D-Fire has become a reference point for the fire management vision community, enabling side-by-side comparison of methods under reproducible and challenging conditions (Boroujeni et al., 17 Mar 2025, Joshi et al., 11 Oct 2025). It supports iterative algorithmic development, robust ablation studies, and the principled evaluation of uncertainty estimation, post-detection rescoring, and compact model deployment.

Subsequent research may focus on integrating D-Fire with multi-modal datasets (e.g., including thermal, NIR imagery), expanding annotation granularity, or leveraging its structure for domain adaptation tasks, ultimately aiming to further reduce the time-to-detection and false alarm rates in operational fire and smoke monitoring systems.