This is my entry for this competition: https://zindi.africa/competitions/arm-unicef-disaster-vulnerability-challenge
This is a fairly straightforward object detection and classification task, here using YOLOv8. That already does a lot of training data augmentation for us.
I have explored the augmentation options and some key hyperparameters to try and get optimal results. Of particular importance are:
- conf: Confidence threshold; trade-off between false negatives and false positives.
- iou: Intersection over Union threshold for Non-Maximal Suppression, i.e. rejecting lower-confidence detections which overlap a stronger detection to some degree. As houses don't overlap (unlike some other probems where we may have a dog behind a cat say), this is set fairly low.