iou(pred_box, truth_box) or iou(prior_box， truth_box) should be calculated here?

[WormCoder]

gluon-cv/gluoncv/model_zoo/yolo/yolo_target.py

Line 193 in 9669d72

batch_ious = self._batch_iou(box_preds, gt_boxes) # (B, N, M)

In the yolov3's paper, ignore_threshold if for the iou between prior(or anchor) box and truth box, i.e. the iou(prior_box， truth_box), so why do you use iou(pred_box, truth_box) here?

[zhreshold]

The order is swapped by transpose. The reason to use it in reverse order is because anchors are of batch size 1, and by doing this we can save batch_size times the computation.