Assess whether detections purely in the sky are still an issue

[fe51]

Pitch

In the past, we have encountered false positives that regularly originated from detections purely in the sky (usually clouds or airplane contrails).

What has to be done

From a dataset provided with detections from various Pyronear stations

• Review some existing models/algos able to detect the outline of the sky
• Identify a relevant one (by testing on images)
• Then, evaluate the number of detections in the dataset that are completely in the sky, those that are partially in the sky (and see which ones are false positives)

Based on the results obtained here and the fact that the false positive rate in the sky is still problematic, we will reuse the work done here to implement a solution that avoids any detection purely in the sky, thereby reducing false positives.

Thanks a lot for you help, happy to discuss it !