Classification or detection ?

[SimJeg]

Hello,

I noticed Pyronear moved from a ResNext model for image classification to a YOLOv5 model for image detection (at least for training !). Similarly, I mainly found two ways to deal with early wildfire detection in the scientific literature :

• use a classifier to classify image patches in which there is smoke, for instance Govil et al., 2020, Fernandes et al., 2022, Dewangan et al., 2022
• use a detection to directly detect smoke on the full image, for instance Yazdi et al., 2022

Why did you choose to move from classification to detection ? Do you have experiments showing detection models are more accurate or is it a matter of runtime / accuracy tradeoff ?

Thanks,
Simon