Which boxes do we use to calc mAP

vision
1

Hi, I’m student and my teacher told me to present how to calculate mAP. After read many topic and some repo, I saw that there are 2 way to chose boxes (if I don’t missunderstood) to calculate mAP.

Let assume that model predicted two tensor as below:

boxes : shape [nbox, 4]                      # coordinate of bouding box
conf    : shape [nbox, num_class]      # confidence score for num_class each box

In the first way:

box_to_calc_mAP = []
conf, labels = torch.max(conf, dim=1)
for cur_class in range(num_class):
   mask = (labels == cur_class)
   cur_boxes = boxes[mask]
   cur_conf    = conf[mask]
   box_to_calc_mAP.append(non_max_suppression(cur_boxes, cur_conf))

mAP = calc_mAP(box_to_calc_mAP)

But the second way:

box_to_calc_mAP = []
for cur_class in range(num_class):
   box_to_calc_mAP.append(non_max_suppression(boxes, conf[:, cur_class]))

mAP = calc_mAP(box_to_calc_mAP)

I feel that the first way is correct, but I saw a repo use the second way, so which is actually true ?