How to match predicted bounding boxes to ground-truth bounding boxes

Question

I have finetuned "fasterrcnn_resnet50_fpn" model from PyTorch for an object detection task, the model has some False Negatives and therefore in some images in the validation dataset the number of the predicted bboxes is less than the number of the ground-truth bboxes. In this case, how to match each predicted bbox to its corresponding ground-truth bbox?

Answer 1

Usually you do the matching by checking if the IoU between the predicted box and GT box is above a given threshold. There is also some Non Max Suppression to deal with multiples predicted boxes corresponding to one true box.