'Collision' resolution for precision in object detection

Question

For object detection we often use metrics based on precision/recall. My question is what is generally the process of matching the prediction and ground truth bound boxes, when there are multiple intersecting boxes.

I.e. consider the image bellow for single-class detection. Red are for the two ground truth boxes, blue and green are predictions. Apparently the blue prediction had higher IoU for both boxes, but as it matches with the left one, can the green one consider the correct prediction, given low enough IoU threshold?

Answer 1

Ok, I'm going to answer my own question.

The basic procedure is basically as follows

1. Calculate all IoU for pairs prediction vs ground box. Leave only the pairs with IoU greater than the threshold
2. If for any prediction there are many ground boxes, leave a single one with maximum IoU
3. If for any ground box there are many predictions after step 2, then the first one counts as True Positive and the rest are False Positive

Another thing is that IoU thresholds are canonically bigger the 0.5, so it's very rare we have to do something on step 2.

Given the picture in the question, the green box would probably have an IoU below a threshold, but otherwise, it's a valid prediction.