faster-rcnn.pytorch icon indicating copy to clipboard operation
faster-rcnn.pytorch copied to clipboard

Zero mAP when testing

Open rkruegs123 opened this issue 5 years ago • 3 comments

Similar to #671 I am getting all zeros for test results, but there is no answer as to why in that post. Like this previous post, it seems as though i am getting non-zero mAP from manual insepction and from the training loss.

Can anyone explain how they might have fixed it?

rkruegs123 avatar Nov 06 '19 02:11 rkruegs123

hey @rkruegs123 , any update on this? I too am facing this issue. The training loss decreases well, but when I run test_net.py I get 0 mAP for all classes

Karthik-Suresh93 avatar Jun 27 '20 08:06 Karthik-Suresh93

I met the same issue , have you solved it ?

Kuuuo avatar Oct 21 '20 01:10 Kuuuo

Do your labels have space or upper cases? In pascal_voc.py, cls = self._class_to_ind[obj.find('name').text.lower().strip()] This code causes your labels mismatch to classes. Maybe you should modify 2 places:

  1. self._classes = (...) Make sure classes are same as your labels.
  2. cls = self._class_to_ind[obj.find('name').text]

JiashengHong avatar Jan 03 '21 05:01 JiashengHong