How can mAP calculation work with sorted labels?

[Sharon507]

I'm confused about how evaluate.py can work correctly when the labels are sorted alphabetically on line 32. The class indices predicted by the model will correspond to the original order of the labels, so can't be compared to the indices of the sorted labels.

I'm working with some COCO images and the APs are coming out as mostly 0s. However, if I remove the command to sort the labels, the AP calculation seems to work ok.

What's confusing me is that this seems to be such a major bug that I can't understand how it wouldn't have been spotted before. Can anyone tell me if I'm missing something? Thank you.

[Sharon507]

I think I now understand this better. Because the labels are also sorted when they are returned from create_training_instances in train.py, this will not be an issue for models that are trained using this code. It's only a problem if evaluate.py is run on a pretrained model that doesn't use labels that have been sorted alphabetically (e.g. a model that uses the COCO labels).

[459below]

I'm glad others came across this as well. Additionally the order of numerical labels is also strictly numerical, i.e. "100" comes before "11".

Given labels:   [u'0', u'1', u'2', u'3', u'4', u'5', u'6', u'7', u'8', u'9', u'10', u'11', u'12', u'13', u'14', u'15', u'16', u'17', u'18', u'19', u'20', u'21', u'23', u'24', u'25', u'26', u'27', u'28', u'29', u'30', u'33', u'34', u'35', u'36', u'37', u'38', u'39', u'40', u'41', u'100']

Training on:    [u'0', u'1', u'10', u'100', u'11', u'12', u'13', u'14', u'15', u'16', u'17', u'18', u'19', u'2', u'20', u'21', u'23', u'24', u'25', u'26', u'27', u'28', u'29', u'3', u'30', u'33', u'34', u'35', u'36', u'37', u'38', u'39', u'4', u'40', u'41', u'5', u'6', u'7', u'8', u'9']

This is incredibly confusing. I'll see if I can fix that somehow. 🤔

[459below]

OK, so adding

labels = sorted(labels)

at

keras-yolo3/utils/bbox.py

Line 59 in 768c524

def draw_boxes(image, boxes, labels, obj_thresh, quiet=True):

shows the same labels in prediction as they were trained. I'll make a PR when I find the time to.