scores in GlobalWheat-WILDS

[waskita]

I'm working with GlobalWheat dataset.
I noticed that evaluation script for GlobalWheat only use boxes with scores > 0.5
It's defined in function DetectionAccuracy
Some models (like YOLOv5) have better scores if we include boxes with confidence less than 0.5
Workaround without modifying the evaluation code:

• remove boxes with small confidence, so it will not be evaluated
• scores for boxes are scaled from 0 ~ 1.0 to 0.5 ~ 1