Training a Fire Detector

This folder illustrate the steps for training YOLOv3 and YOLOv3-tiny to detect fire in images and videos.

First a fire dataset of labeled images is collected from internet. The images with their annotations have been prepared and converted into YOLO format and put into one folder to gather all the data. Therefore, the data folder contains images ('*jpg') and their associated annotations files ('.txt') with the same name.

The annotations need to be converted into YOLO format, which is :

<class_id, x_c/W, y_c/H, h/H, w/W>

AI-lab

Launch AI-LAB

Pull AI-lab from Docker Hub

docker pull aminehy/ai-lab

Run the AI-lab and start your development

xhost +

then

docker run -it --rm
--runtime=nvidia
-v $(pwd):/workspace \
-w /workspace \
-v /tmp/.X11-unix:/tmp/.X11-unix \
-e DISPLAY=$DISPLAY \
-p 8888:8888 -p 6006:6006 aminehy/ai-lab:latest

Retrain YOLO on a Costum Dataset

YOLOv3

1. Create a customized configuration file for YOLO model from cfg/yolov3.cfg

   cp cfg/yolov3.cfg yolov3-obj.cfg
   

   yolov3-obj.cfg
       [net]
       # Testing
       # batch=64
       # subdivisions=8
       # Training
       batch=64
       subdivisions=16
       width=416
       height=416
       channels=3
       momentum=0.9
       decay=0.0005
       angle=0
       saturation = 1.5
       exposure = 1.5
       hue=.1
       ...
   

2. Open yolov3-obj.cfg and edit its content with the appropriate information. In this application we have only one object to detect, 'fire', thus, nb_class = 1

   * batch = 64
   * subdivision=8 (increase if `Out of memory`)
   * filters = (nb_class+5)*3: filters = 255 => filter = 18
   * classes = nb_class: classes = 80  => classes = 1
   * max_batches=classes*2000: max_batches = 2000
   * steps=80% and 90% of max_batches : steps = 1600, 1800
   
   

3. Create a file train.txt that lists paths of all annotated images (*.jpg) of the dataset data/obj/

   train.txt
       data/obj/pic (132).jpg
       data/obj/img (42).jpg
       data/obj/img (90).jpg
       data/obj/pic (148).jpg
       data/obj/small (82).jpg
       data/obj/small (55).jpg
       data/obj/small (73).jpg
       data/obj/small (53).jpg
       data/obj/pic (42).jpg
       ...
   
   

4. Create obj.names and list the classes names

   obj.names
       fire
   

5. Create and setup a data file obj.data

   obj.data
       classes = 1
       train  = train.txt
       valid  = train.txt
       names = obj.names
       backup = backup/
   

6. Download pre-trained weights for the convolutional layers (154 MB): https://pjreddie.com/media/files/darknet53.conv.74

7. Start training by using the command line

   ./darknet detector train yolov3-obj.cfg obj.data darknet53.conv.74
   

8. test:

./darknet detector test obj.data  yolov3-obj.cfg backup/yolov3-obj_final.weights

YOLOv3-tiny

1. Create a customized configuration file for YOLO model from cfg/yolov3-tiny_obj.cfg

   cp cfg/yolov3.cfg yolov3-obj.cfg
   

   yolov3-obj.cfg
       [net]
       # Testing
       # batch=64
       # subdivisions=8
       # Training
       batch=64
       subdivisions=16
       width=416
       height=416
       channels=3
       momentum=0.9
       decay=0.0005
       angle=0
       saturation = 1.5
       exposure = 1.5
       hue=.1
       ...
   

2. Open yolov3-obj.cfg and edit its content with the appropriate information. In this application we have only one object to detect, 'fire', thus, nb_class = 1

   * batch = 64
   * subdivision=8 (increase if `Out of memory`)
   * filters = (nb_class+5)*3: filters = 255 => filter = 18
   * classes = nb_class: classes = 80  => classes = 1
   * max_batches=classes*2000: max_batches = 2000
   * steps=80% and 90% of max_batches : steps = 1600, 1800
   
   

3. Get pre-trained weights yolov3-tiny.conv.15

./darknet partial cfg/yolov3-tiny.cfg fire/model/yolov3-tiny.weights fire/model/yolov3-tiny.conv.15 15

4. Training

./darknet detector train obj.data yolov3-tiny-obj.cfg fire/model/yolov3-tiny.conv.15

6. Results

v3 (mse loss, Normalizer: (iou: 0.750000, cls: 1.000000) Region 23 Avg (IOU: 0.664938, GIOU: 0.652729), Class: 0.999684, Obj: 0.376908, No Obj: 0.000314, .5R: 1.000000, .75R: 0.000000, count: 3

 2000: 0.604868, 0.560218 avg loss, 0.000010 rate, 1.346011 seconds, 128000 images
Saving weights to backup//yolov3-tiny-obj_2000.weights
Saving weights to backup//yolov3-tiny-obj_last.weights
Saving weights to backup//yolov3-tiny-obj_final.weights

Test of the Trained Fire Detection Application

(Note: You might want to recompile the DarkNet on your computer: Edit the Makefile.txt then run make)

Image

Detect fire in an imahge file

```
./darknet detector test fire/data/obj.data fire/cfg/yolov3-tiny-obj.cfg fire/model/yolov3-tiny-obj_final.weights fire/data/obj/img (9).jpg
```

Video

• Detect fire in real-time video stream from webcam

  ./darknet detector demo fire/data/obj.data fire/cfg/yolov3-tiny-obj.cfg fire/model/yolov3-tiny-obj_final.weights
  

  Note: add the option to save the output video: -out_filename filename.mp4

• Test on video stream from file

  ./darknet detector demo fire/data/obj.data fire/cfg/yolov3-tiny-obj.cfg fire/model/yolov3-tiny-obj_final.weights fire/videos/test.mp4
  

• Example:

  ./darknet detector demo fire/data/obj.data fire/cfg/yolov3-tiny-obj.cfg fire/model/yolov3-tiny-obj_final.weights fire/videos/DJI_0070.MOV -out_filename ./fire/videos/DJI_0070_output.mp4

Pre-trained Models

There are weights-file for different cfg-files (smaller size -> faster speed & lower accuracy:

• yolov3-openimages.cfg (247 MB COCO Yolo v3) - requires 4 GB GPU-RAM: https://pjreddie.com/media/files/yolov3-openimages.weights

• yolov3-spp.cfg (240 MB COCO Yolo v3) - requires 4 GB GPU-RAM: https://pjreddie.com/media/files/yolov3-spp.weights

• yolov3.cfg (236 MB COCO Yolo v3) - requires 4 GB GPU-RAM: https://pjreddie.com/media/files/yolov3.weights

• yolov3-tiny.cfg (34 MB COCO Yolo v3 tiny) - requires 1 GB GPU-RAM: https://pjreddie.com/media/files/yolov3-tiny.weights

• enet-coco.cfg (EfficientNetb0-Yolo- 45.5% mAP@0.5 - 3.7 BFlops)

enetb0-coco_final.weights and yolov3-tiny-prn.cfg (33.1% mAP@0.5 - 3.5 BFlops - more)

Reference

• https://github.com/AlexeyAB/darknet