Note: Our project consists of two sub-projects: one is ¹Benchmarking Multi-Scene Fire and Smoke Detection and the other is ²Fire and Smoke Detection with Burning Intensity Representation. Our FSD Code release time will be after Dec. 30.

¹Benchmarking Multi-Scene Fire and Smoke Detection

Xiaoyi Han¹, Nan Pu², Zunlei Feng¹, Yijun Bei¹, Qifei Zhang¹, Lechao Cheng^3,✉, Liang Xue⁴,

¹School of Software Technology, Zhejiang University, ²University of Trento,
³Hefei University of Technology, ⁴Suzhou City University
^✉Corresponding Author
Accepted to PRCV 2024

arxiv Previous FSD Datasets MS-FSDB Code Springer

Note

Dear Visitor,
Hello! If you would like to use our Unified FSD Datasets, please click on the Google Drive link and provide your institution (e.g., school, company, or None if applicable) along with a brief description of your intended use for the datasets. After that, please send a request for access to the shared folder. Note: Applications without an institution and a description of the intended use will be rejected.
Additionally, our code includes data processing and evaluation methods. You can view it on our GitHub. Thank you!

Best regards,

Xiaoyi Han

Abstract

The current irregularities in existing public Fire and Smoke Detection (FSD) datasets have become a bottleneck in the advancement of FSD technology. Upon in-depth analysis, we identify the core issue as the lack of standardized dataset construction, uniform evaluation systems, and clear performance benchmarks. To address this issue and drive innovation in FSD technology, we systematically gather diverse resources from public sources to create a more comprehensive and refined FSD benchmark. Additionally, recognizing the inadequate coverage of existing dataset scenes, we strategically expand scenes, relabel, and standardize existing public FSD datasets to ensure accuracy and consistency. We aim to establish a standardized, realistic, unified, and efficient FSD research platform that mirrors real-life scenes closely. Through our efforts, we aim to provide robust support for the breakthrough and development of FSD technology.

Our MS-FSDB and other FSD Datasets Overview

At first, we propose a new Multi-Scene Fire and Smoke Detection Benchmark (MS-FSDB) comprising 12,586 images, depicting 2,731 scenes as illustrated in the aforementioned images. Most images within our benchmark possess dimensions exceeding 600 pixels in either length or width. Unlike previous public Fire and Smoke Detection (FSD) datasets, our benchmark not only includes flame detection but also smoke detection tasks. Additionally, it captures complex scenes featuring occlusion, multiple targets, and various viewpoints.

To access MS-FSDB, please click here.

Then, we compared our benchmark with the most prevalent and easily accessible Fire and Smoke Detection (FSD) datasets, subjecting these popular datasets to the same secondary processing as our own, including labeling and image selection, to facilitate research in the field of fire detection. These datasets include Fire-Smoke-Dataset¹, Furg-Fire-Dataset², VisiFire³, FIRESENSE⁴, BoWFire⁵.

To access Public FSD Datasets, please click here.

To access Fire-Smoke-Dataset, please click here.
1.DeepQuestAI, Fire-smoke-dataset (2021). URL https://github.com/DeepQuestAI/Fire-Smoke-Dataset.

To access Furg-Fire-Dataset, please click here.

2.V. H¨uttner, C. R. Steffens, S. S. da Costa Botelho, First response fire combat: Deep leaning based visible fire detection, in: 2017 Latin American Robotics Symposium (LARS) and 2017 Brazilian Symposium on Robotics (SBR), IEEE, 2017, pp. 1–6. doi:10.1109/SBR-LARS-R.2017.8215312.

To access VisiFire, please click here.

3.B. U. Toreyin, A. E. Cetin, Online detection of fire in video, in: 2007 IEEE Conference on Computer Vision and Pattern Recognition, 2007, pp. 1–5. doi: 10.1109/CVPR.2007.383442.

To access FIRESENSE, please click here.

4.K. Dimitropoulos, P. Barmpoutis, N. Grammalidis, Spatio-temporal flame modeling and dynamic texture analysis for automatic video-based fire detection, IEEE Transactions on Circuits and Systems for Video Technology 25 (2) (2015) 339–351. doi:10.1109/TCSVT.2014.2339592.

To access BowFire, please click here.

5.D. Y. T. Chino, L. P. S. Avalhais, J. F. Rodrigues, A. J. M. Traina, Bowfire: Detection of fire in still images by integrating pixel color and texture analysis, in: 2015 28th SIBGRAPI Conference on Graphics, Patterns and Images, 2015, pp. 95–102. doi:10.1109/SIBGRAPI.2015.19.

Benchmark Presentation and Experimental Results

Multi-Scene in FSDB: residential building, air crush, indoor clutter, electric power station, etc.

There is Illustrations of several Fire and Smoke Detection (FSD) datasets statistics. (a) represents the statistics comparison of several FSD datasets scenes. (b) provides more detailed statistics for our benchmark. In (b), the area of the sector corresponds to the number of scenes. Additionally, the total number of scenes equals 2731, as displayed in (a).

Baseline models comparison across different datasets, where Fire represents the Average Precision (%) of fire, Smoke represents the Average Precision (%) of smoke, and mAP represents the mean Average Precision (%) of fire and smoke. “s” represents the input image of small size, while “l” represents the input image of large size, and the specific content is in the subsection “Basic Experiments”. F-RCNN means Faster RCNN. Ours means our method in ²Fire and Smoke Detection with Burning Intensity Representation.

Cross experimental results in (a), where results represent the mean Average Precision (%) of fire and smoke. the table on the right, we train baseline models using the proposed miniMS-FSDB and then test with previous FSD dataset. For example, a mAP of 45.8 represents the result of the SSD model trains on our miniMS-FSDB and then test using the Fire-Smoke-Dataset. The table on the left, we train baseline models using previous FSD dataset and then test with miniMS-FSDB. For example, a mAP of 45.5 represents the result of the SSD model trains on Fire-Smoke-Dataset and the test using our miniMS-FSDB. We uniformly choose small input images for all baseline models to ensure consistency and comparability in experiments. F-RCNN means Faster RCNN and Ours means our method in ²Fire and Smoke Detection with Burning Intensity Representation.
We train baselines on different FSD datasets and then test with the new testing set in (b), where mAP represents the mean Average Precision (%) of fire and smoke. We use the input image of small size. F-RCNN means Faster RCNN,and Ours means our method in ²Fire and Smoke Detection with Burning Intensity Representation.

BibTeX

@InProceedings{hanprcvfsd,
      author="Han, Xiaoyi and Pu, Nan and Feng, Zunlei and Bei, Yijun and Zhang, Qifei and Cheng, Lechao and Xue, Liang", 
      editor="Lin, Zhouchen and Cheng, Ming-Ming and He, Ran and Ubul, Kurban and Silamu, Wushouer and Zha, Hongbin and Zhou, Jie and Liu, Cheng-Lin", 
      title="Benchmarking Multi-Scene Fire and Smoke Detection",booktitle="Pattern Recognition and Computer Vision", 
      year="2025", publisher="Springer Nature Singapore", address="Singapore", pages="203--218", 
      abstract="The current irregularities in existing public Fire and Smoke Detection (FSD) datasets have become a bottleneck in the advancement of FSD technology. Upon in-depth analysis, we identify the core issue as the lack of standardized dataset construction, uniform evaluation systems, and clear performance benchmarks. To address this issue and drive innovation in FSD technology, we systematically gather diverse resources from public sources to create a more comprehensive and refined FSD benchmark. Additionally, recognizing the inadequate coverage of existing dataset scenes, we strategically expand scenes, relabel, and standardize existing public FSD datasets to ensure accuracy and consistency. We aim to establish a standardized, realistic, unified, and efficient FSD research platform that mirrors real-life scenes closely. Through our efforts, we aim to provide robust support for the breakthrough and development of FSD technology. The project is available at https://xiaoyihan6.github.io/FSD/.",
isbn="978-981-97-8795-1"}

******************************************************************************************************************************************************************************************************************************************************************************************************************************************************************

²Fire and Smoke Detection with Burning Intensity Representation

Xiaoyi Han¹, Yanfei Wu², Nan Pu³, Zunlei Feng¹, Qifei Zhang¹, Yijun Bei¹, LeChao Cheng^4,✉,

¹School of Software Technology, Zhejiang University, ²China Mobile (Suzhou) Software Technology Co., Ltd.,
³University of Trento, ⁴Hefei University of Technology
^✉Corresponding Author
Accepted to ACM MM asia 2024

arxiv Code ACM

Abstract

An effective Fire and Smoke Detection (FSD) and analysis system is of paramount importance due to the profound destructive potential of fire disasters. However, many existing FSD methods directly employ generic object detection techniques without accounting for the transparency of the fire and smoke, inevitably leading to imprecise localization of fire and smoke areas and consequently diminishing detection performance of fire and smoke. To address this issue, a new Attentive Transparency Detection Head (ATDH) is proposed to improve the accuracy of transparent target detection while retaining the robust feature extraction and fusion capabilities of conventional detection algorithms. In addition, Burning Intensity (BI) is introduced as a pivotal feature for fire-related downstream risk assessment in traditional FSD methodologies. Extensive experiments on multiple FSD datasets showcase the effectiveness and versatility of the proposed FSD model.

Method and Experimental Results

Our Fire and Smoke Detection Method, which is called Attentive Fire Smoke Detection Model (a-FSDM). And it consists of three main components: Feature Extraction, Feature Fusion Grouping, and Attentive Transparency Detection Head (ATDH) of FSD task. Feature Fusion Grouping is used to fuse semantic and spatial information from various convolutional layers to prevent information loss. The ATDH is responsible for classification, regression, and centerness.

If you want to compare baseline models across different datasets, you can refer to ¹Benchmarking Multi-Scene Fire and Smoke Detection.
And then, comparison between generic detection heads and the Attention Transparency Detection Head (ATDH) across the MS-FSDB is provided in (a), where Fire, Smoke, and mAP are discussed in the subsection 'Setting and Details.' “s” represents the input image of small size, while “l” represents the input image of large size.

In (1), the attention mechanism algorithm is added to FCOS on the MS-FSDB, where Fire, Smoke, and mAP are discussed in the subsection 'Setting and Details.' The input image of small size is used.
In (2), the evidence supporting the efficacy of the ATDH, (a) original image, (b) CAM visualization results of FCOS with SENet [13], (c) CAM visualization results of FCOS with SKNet [22]. (d) CAM visualization results of FCOS with ATDH.

The performance of different models is shown in detail, covering seven detection models in terms of key metrics such as detection accuracy, mAP, performance (Gflops), and the number of parameters (M) for each category. Fire, Smoke, and mAP are discussed in the subsection 'Setting and Details.' We use small-sized input images from the miniMS-FSDB. In the models, [17] represents YOLOv5, [18] stands for YOLOv8, [25] denotes SSD, [24] signifies RetinaNet, [33] means Faster R-CNN, [37] refers to FCOS, and 'ours' represents the proposed a-FSDM model. Additionally, 's', 'x', 'n', 'l', and 'm' represent different versions of the models.

BibTeX

@article{han2024mmasia,
  title={Fire and Smoke Detection with Burning Intensity Representation},
  author={Xiaoyi Han and Yanfei Wu and Nan Pu and Zunlei Feng and Qifei Zhang and Yijun Bei and Lechao Cheng},
  booktitle={The 6th ACM Multimedia Asia conference},
  year={2024}
}

1Benchmarking Multi-Scene Fire and Smoke Detection

Note

Abstract

Our MS-FSDB and other FSD Datasets Overview

Benchmark Presentation and Experimental Results

Multi-Scene in FSDB: residential building, air crush, indoor clutter, electric power station, etc.

BibTeX

2Fire and Smoke Detection with Burning Intensity Representation

Abstract

Method and Experimental Results

BibTeX

¹Benchmarking Multi-Scene Fire and Smoke Detection

²Fire and Smoke Detection with Burning Intensity Representation