WildFireSmokeDetection

Submission for hackathon conducted by AI For Mankind (https://aiformankind.org/) to early detect WildFire smoke . See details of hackathon at https://aiformankind.org/lets-stop-wildfires-hackathon-2.0/

Thanks to AIForManKind for providing Quick Start Demo https://github.com/aiformankind/wildfire-smoke-detection-camera and providning label Image Data Set.

Saved Model

Submitted fined tune model is trained with EfficientDet-d3 using TensorFlow.

Data Set - 737 images. After augmenting (Horizontal Flip and added brightness), dataset was :-

Training Images : 1739

Validation Images : 111

Total training steps : 107000

Fine Tuned Model

Saved model can be downloaded from https://drive.google.com/drive/folders/1R54ZCvD9-aNc-q59ZxUK_go9wO5qJKku?usp=sharingv

How to do Training and Inference

See Model Training notebook to do train youe model on smoke images.

For doing inference from saved model refer to inference notebook

Resources

WildFire Resources

• FUEGO Wildfire Detection Slides by Kinshuk Govil
• [Wildland Fire Assessment System] (https://journals.sagepub.com/doi/pdf/10.1155/2014/597368)
• [How Wildfire Works] (https://science.howstuffworks.com/nature/natural-disasters/wildfire.htm/printable)

Tensorflow Resources

• Tensorflow Quickstart
• TF Objection Detection API
• [Object detection inference] (https://github.com/tensorflow/models/blob/master/research/object_detection/colab_tutorials/inference_from_saved_model_tf2_colab.ipynb)

Other Resources

• Faster RCNN ResNnet
• Train EfficientDet in TensorFlow
• Data Augmentation using roboflow
• Train object detection with Keras

Others model Tried

• YOLOV5 - With YOLOV5 with our dataset full smoke images were not detected properly.

• SSD Mobile to solve this problem, but in results we found some limitations with some pattern of images. Training was very slow

• FatserRCNN ResNet101 - Got very best accuracy and lowest loss with this. But it was giving many False Positive for Fog images test.

• Faster_rcnn_inception_resnet_v2_atrous_coco gives good results for true positives but the prediction time is very high and do not solve False Postives problem(predicting fog as smoke)

• Segmentation part of this problem is also tried with Detectron2 model by preparing data from Labelme and then converted it to COCO with labelme2coco.py.

• AP factor for segmenatation part was very less, so we are not including with our results.

Results