This project is focused on creating a fire detection model to enhance rescue efficiency with improved accuracy and speed, aiming for real-time processing.
Project inspired by: real-time-fire-segmentation-deep-learning
Our goal was to reproduce the experiment results and test the model on random YouTube videos to evaluate accuracy.
Result 2 - Random Video from YouTube
The encoder consists of a DCNN with 16 convolution layers and an ASPP module (Atrous Spatial Pyramid Pooling). The activation function is a combination of both ReLU and HardSwish. The Lion optimizer and Focal Loss are used.
Loss, MPA [Mean Pixel Average], and MIoU [Mean Intersection over Union] recorder after every epoch- Clone the repository.
- Download the dataset files "Images for fire segmentation" and "Masks annotation for fire segmentation" from IEEE Dataport and move them to the 'data' folder.
- Follow the instructions in the
main.ipynb
file.