in ocr/utils/iou_loss.py [0:0]
def hybrid_forward(self, F, pred, label, sample_weight=None):
'''
Interpreted from: https://www.pyimagesearch.com/2016/11/07/intersection-over-union-iou-for-object-detection/
Steps to calculate IOU
1) Calculate the area of the predicted and actual bounding boxes
2) Calculate the area of the intersection between the predicting and actual bounding box
3) Calculate the log IOU by: log(intersection area / (union area))
3) If the bounding boxes do not overlap with one another, set the iou to zero
4) Calculate the negative mean of the IOU
'''
pred_area = pred[:, 2] * pred[:, 3]
label_area = label[:, 2] * label[:, 3]
x1_1, y1_1 = pred[:, 0], pred[:, 1]
x2_1, y2_1 = pred[:, 0] + pred[:, 2], pred[:, 1] + pred[:, 3]
x1_2, y1_2 = label[:, 0], label[:, 1]
x2_2, y2_2 = label[:, 0] + label[:, 2], label[:, 1] + label[:, 3]
x_overlaps = F.logical_or(x2_1 < x1_2, x1_1 > x2_2)
y_overlaps = F.logical_or(y2_1 < y1_2, y1_1 > y2_2)
overlaps = F.logical_not(F.logical_or(x_overlaps, y_overlaps))
x1_1 = x1_1.expand_dims(0)
y1_1 = y1_1.expand_dims(0)
x2_1 = x2_1.expand_dims(0)
y2_1 = y2_1.expand_dims(0)
x1_2 = x1_2.expand_dims(0)
y1_2 = y1_2.expand_dims(0)
x2_2 = x2_2.expand_dims(0)
y2_2 = y2_2.expand_dims(0)
x_a = F.max(F.concat(x1_1, x1_2, dim=0), axis=0)
y_a = F.max(F.concat(y1_1, y1_2, dim=0), axis=0)
x_b = F.min(F.concat(x2_1, x2_2, dim=0), axis=0)
y_b = F.min(F.concat(y2_1, y2_2, dim=0), axis=0)
inter_area = (x_b - x_a) * (y_b - y_a)
iou = F.log(inter_area) - F.log(pred_area + label_area - inter_area)
loss = gluon.loss._apply_weighting(F, iou, self._weight, sample_weight)
loss = F.where(F.logical_not(overlaps), F.zeros(shape=overlaps.shape), loss)
mean_loss = F.mean(loss, axis=self._batch_axis, exclude=True)
return -mean_loss