A running version of the application is hosted on Huggingface.
The proposed system consists of two main solutions. In the first solution we design, train, and evaluate a custom CNN. Regarding the second solution, we will fine-tune a pretrained vision transformer, the Google ViT [1], hosted on the Hugging Face Hub [2]. Results are logged and proposed through WanDB [3]. The performance of the two solutions will be compared and different aspects will be discussed.
The framework used for this experiment is PyTorch with the PyTorch Lightning Module. The final web application to showcase the model is written in Python using the Streamlit library and hosted on the Hugging Face Hub.
The dataset used in this project is self-built by merging two datasets from Kaggle. In particular, we used samples from ”train_fire” and ”train_smoke” from [4], and all the samples (mixed together from further splitting) from [5].
The final dataset consists of 2525 samples for each of the three classification classes (fire, smoke, normal). This dataset will be then split into train, validation, and test sets. The images in the dataset depict fires of different magnitude, from different perspectives (aerial, far, near) thus simulating images from different devices, such as webcams and drones.
To showcase the final model, we developed a simple web application hosted on Huggingface. The demo allows the user to upload an image and predict if there is a smoke, fire or normal situation. We report here a screenshot of the GUI.
[1] B. Wu et al., Visual Transformers: Token-based Image Representation and Processing for Computer Vision. arXiv, 2020. doi: 10.48550/ARXIV.2006.03677.
[2] Hugging face[EB/OL]. https://huggingface.co.
[3] Biewald L. Experiment tracking with weights and biases. Software available from wandb.com.
[4] KUTLU K. Forest fire dataset on kaggle[EB/OL]. https://www.kaggle.com/datasets/kutaykutlu/forest-fire?select=train_fire.
[5] PRASAD M S. Forest fire images dataset on kaggle[EB/OL]. https://www.kaggle.com/datasets/mohnishsaiprasad/forest-fire-images.