MSFFD

English | 中文简体

1. MSFFD Download Link

We provide two ways to access the MSFFD:

• Roboflow, the address of this dataset is:https://universe.roboflow.com/yunnan-university/synthetic-fire-smoke

• Cloud Drive, we provide the data in different annotation formats for different target detection algorithms (coco, darknet, yolov6, yolov7, yolov8).

  • The download links for Google Cloud Drive are: coco, darknet, yolov6, yolov7, yolov8, different_weather_and_time:.

  • The download links for baidu cloud disk are:

    • coco: https://pan.baidu.com/s/18w33GJtQar0gfOM3ocuR1A?pwd=yqv8 提取码：yqv8

    • darknet: https://pan.baidu.com/s/1jqzTn93ixUf_TXhrCRaW4w?pwd=c8dx 提取码：c8dx

    • yolov6: https://pan.baidu.com/s/1X_2mmSyAMCshnYJIOeQnYg?pwd=fk3r 提取码：fk3r

    • yolov7: https://pan.baidu.com/s/1HgC5YpEtF3VJgLt5zRD_Uw?pwd=q89a 提取码：q89a

    • yolov8: https://pan.baidu.com/s/1VDbQBMD3SamF0a7_r9aLow?pwd=2rsq 提取码：2rsq

    • different_weather_and_time: https://pan.baidu.com/s/1O4QKkeJubQa6lZKRdnwR6w?pwd=txil 提取码：txil

2. Scenario introduction

In order to achieve a realistic simulation of forest fires, we use Unreal Engine 5 to build a diverse forest scene.

Unreal Engine 5 can meet the details of real-world forest terrain, weather, time, lighting, texture, etc. It can also simulate multi-scale forest fires in different periods, which is ideal for the construction of synthetic datasets.

We built eight scenes with different terrain through Unreal Engine 5, containing three types of weather conditions: sunny, foggy, and rainy-snowy, and three different times of day, evening, and night.

The following table shows the details of the raw data.

	Scenario	Resolution	FPS	Duration	Size	Weather	The number of objects	Description
1	a	1480×684	17.6	126 seconds	270.7 Mb	Sunny	Multiple Objects	Daylight, mountains, and plains
2	a	1480×684	17.7	85 seconds	227.5 Mb	Sunny	No Objects	Evening, mountains, and plains
3	a	1480×684	17.2	109 seconds	30.2 Mb	Sunny	Multiple Objects	Night, mountains, and plains
4	a	1480×684	13.8	79 seconds	184.1 Mb	Sunny	No Objects	Evening, mountains, and plains
5	a	776×452	18.6	113 seconds	75.3 Mb	Sunny	No Objects	Daylight, mountains, and plains
6	a	1480×684	17.9	82 seconds	9.7 Mb	Sunny	Single Object	Night, mountains, and plains
7	b	1480×684	29.9	115 seconds	496.2 Mb	Sunny	Multiple Objects	Daylight, mountains, and plains
8	b	1480×684	28	82 seconds	314.4 Mb	Sunny	No Objects	Daylight, Evening, Night, mountains, and plains
9	b	1480×684	29.9	101 seconds	477.4 Mb	Sunny	Single Object	Daylight, mountains, and plains
10	b	1480×684	26.5	83 seconds	43.4 Mb	Sunny	Single Object	Night, mountains, and plains
11	c	1480×684	18.5	195 seconds	258.4 Mb	Sunny	Multiple Objects	Daylight, Night, mountains, and plains
12	c	1480×684	21.2	190 seconds	423.3 Mb	Sunny	Single Object	Daylight, Evening, Night, mountains, and plains
13	d	1480×684	9.99	133 seconds	145.9 Mb	Sunny	Single Object	Daylight, Evening, Night, mountains, lakes, and plains
14	d	1480×684	15.7	122 seconds	169.2 Mb	Sunny	Multiple Objects	Daylight, Evening, Night, mountains, lakes, and plains
15	d	1480×684	11.1	125 seconds	136.5 Mb	Sunny	No Objects	Daylight, Evening, Night, mountains, lakes, and plains
16	e	1280×720	30	148 seconds	348 Mb	Sunny	Multiple Objects	Daylight, mountains, and lakes
17	e	1280×720	30	97 secnds	130.2 Mb	Sunny	Multiple Objects	Night, mountains, and lakes
18	e	1280×720	30	89 seconds	92 Mb	Sunny	Multiple Objects	Evening, mountains, and lakes
19	f	1280×720	24.6	60 seconds	342.8 Mb	Rainy&Fog	Multiple Objects	Daylight, mountains, lakes, and plains
20	f	1280×720	23.7	29 seconds	20.2 Mb	Rainy&Fog	Multiple Objects	Evening, Night, mountains, lakes, and plains
21	f	1280×720	10.6	57 seconds	6.7 Mb	Rainy&Fog	Multiple Objects	Night, mountains, lakes, and plains
22	f	1280×721	10.6	36 seconds	28.1 Mb	Rainy&Fog	Multiple Objects	Evening, mountains, lakes, and plains
23	g	1280×722	10.2	99 seconds	206 Mb	Sunny	Multiple Objects	Daylight, plains
24	g	1280×723	10.5	27 seconds	21.8 Mb	Sunny	Multiple Objects	Night, plains
25	g	1280×724	9.3	61 seconds	92 Mb	Sunny	Multiple Objects	Evening, plains
26	h	1280×724	24.7	40 seconds	251.1 Mb	Snow	Multiple Objects	Daylight, mountains, and plains
27	h	1280×724	26	33 seconds	143.4 Mb	Snow	Multiple Objects	Night, mountains, and plains
28	h	1280×724	25.3	49 seconds	183.6 Mb	Snow	Multiple Objects	Evening, mountains, and plains

2.1 Diverse scenarios

We built eight multi-scale forest scenarios with different terrain and vegetation.

2.2 Multiple Weather

MSFFD simulates not only sunny weather scenes, but also rainy, foggy and snowy weather scenes.

2.3 Multiple times of day

MSFFD considers multimodal forest fire images at different points of the day (day, evening and night).

2.4 Different number of fire objects

In a multimodal forest scenario with multiple terrains, multiple meteorological conditions, and multiple time points, we set up three fire target number scenarios: no objects, single object, and multiple objects, aiming to fully evaluate the performance of the simulated multimodal forest fire dataset in terms of target detection algorithm performance.

3. Details of the dataset

The following table shows the distribution of the images and annotated boxes of the dataset in different weather and at different times of the day.

Multimodal scenarios	Number of images	fire	smoke	Number of annotation boxes	Average number of instances boxes per image
Sunny	3292	8423	7095	15518	4.7
Snowy	312	772	547	1319	4.2
Rainy & Fog	370	432	494	926	2.5
Daytime	2209	4684	3807	8491	3.8
Evening	560	2052	1581	3633	6.5
Night	1205	2891	2748	5639	4.7

The following table shows the size distribution of all comment boxes.

class	small boxes	medium boxes	large boxes	total
fire	4848	4114	665	9627
smoke	1883	2037	4216	8136
all	6731	6151	4881	17763

The following figure shows the distribution of all annotation boxes in the dataset, as shown below.

The distribution of the centroid coordinates of the fire and smoke annotation boxes in different modes is shown in the following figure.

The size distribution of fire and smoke annotation boxes for different modes is shown in the following figure.

4. Experimental results

4.1 Experimental results of MSFFD in single-stage object detection algorithm

Model	#Param.	Flops	Weight Size	AP50:95	AP50	APFire	APSmoke	Precision	Recall	F1-socre
YOLOv8m	25.9M	78.9G	52Mb	51.2	87.3	88.7	86	83.7	86.9	85.3
YOLOv8l	43.7M	165.2G	87.7Mb	51.6	86.8	88.9	84.8	84.6	80.1	82.3
YOLOv8x	68.2M	258.1G	136.7Mb	47.3	87.1	88.5	85.8	84.4	80.2	82.2
YOLOv7	36.5M	103.2G	74.8Mb	48.5	87.5	88.7	86.2	85.3	80.5	82.8
YOLOv7-X	70.8M	188G	142.1Mb	47.3	87.1	88.5	85.8	84.4	80.2	82.2
YOLOv6-M	34.9M	85.8G	74.9Mb	44.5	83.4	84.2	82.6	83.8	80.7	82.2
YOLOv6-L	59.5M	150.7G	117.6Mb	48.8	86.3	87.2	85.4	84.2	79.8	81.9
YOLOv6-M6	79.6M	379.5G	72.51Mb	43.7	82.9	82	85.1	82.6	80	81.3
YOLOv6-L6	140.4M	673.4G	268.4Mb	46.3	83.3	83.9	82.6	83.6	79.7	81.6
YOLOv5m	21.2M	49.0G	42.2Mb	45.6	87.1	88.1	86	86.6	81.5	84.0
YOLOv5l	46.5M	109.1G	92.8Mb	45.9	86.7	87.9	85.4	85.2	81.6	83.4
YOLOv5x	86.7M	205.7G	173.1Mb	45.7	86.3	87.3	85.3	84.6	82.3	83.4
YOLOv4	-	59.6G	256Mb	-	75.3	75.7	74.8	66	73	69.3
YOLOv4-CSP	-	50.3G	210.3Mb	-	81.5	83.8	79.3	78	80	79.0
YOLOv3	-	65.3G	246.3Mb	-	82.3	82.6	82	83	77	79.9
YOLOv3-spp	-	65.7G	250.5Mb	-	82.1	82.3	81.8	83	78	80.4

4.2 Experimental results in two-stage object detection algorithm

Model	Image Size	Weight Size	AP50:95	AP50
Faster R-CNN R50-DC5	640	1300Mb	38.6	78.9
Faster R-CNN R50-FPN	640	333.3Mb	42.1	82.7
Faster R-CNN R50-C4	416	257.5Mb	24.8	59.8
Faster R-CNN R101-C4	416	422Mb	26.9	60
Faster R-CNN R101-DC5	416	1400Mb	25.3	57.8
Faster R-CNN R101-FPN	416	365.2Mb	35.9	76.1
RPN R50	416	103.5Mb	25.4	55.8
Fast R-CNN R50-FPN	416	318.7Mb	28.5	59.8

4.3 Experimental Results of MSFFD in Lightweight objects detection algorithm

Model	#Param.	Flops	Weight Size	AP50:95	AP50	APFire	APSmoke	Precision	Recall	F1-socre
YOLOv8n	3.2M	8.7G	6.3Mb	49.5	86.4	87.1	85.7	83.2	81.2	82.2
YOLOv8s	11.2M	28.6G	22.5Mb	50.3	86.4	87.9	85	83.1	81.4	82.2
YOLOv6-N	4.7M	11.4G	10.0Mb	42	82.2	82.9	81.5	83.3	78.8	81.0
YOLOv6-S	18.5M	45.3G	10.2Mb	44.4	84.1	85	83.2	83.9	80.3	82.1
YOLOv6-N6	10.4M	49.8G	21.83Mb	42.1	82.1	82.8	81.3	83.7	78.9	81.2
YOLOv6-S6	41.4M	198G	85.84Mb	43.2	83.1	83.8	82.5	83.6	79.4	81.4
YOLOv5n	1.9M	4.5G	3.8Mb	40.1	84.9	86.1	83.7	84.3	79.1	81.6
YOLOv5s	7.2M	16.5G	14.4Mb	43.7	86.4	87.9	84.9	83.7	80.9	82.3
YOLOv4-tiny	-	6.8G	23.5Mb	-	64.5	51.9	77.1	80	58	67.2
YOLOv3-tiny	-	5.4G	34.7Mb	-	71.7	69.6	73.9	76	72	73.9

4.4 Visualization of experimental results

                        @misc{
                            synthetic-fire-smoke-w6bgy_dataset,
                            title = { synthetic fire-smoke Dataset },
                            type = { Open Source Dataset },
                            author = { classificationstaff },
                            howpublished = { \url{ https://universe.roboflow.com/classificationstaff/synthetic-fire-smoke-w6bgy } },
                            url = { https://universe.roboflow.com/classificationstaff/synthetic-fire-smoke-w6bgy },
                            journal = { Roboflow Universe },
                            publisher = { Roboflow },
                            year = { 2024 },
                            month = { nov },
                            note = { visited on 2024-12-25 },
                            }