1

I am trying to use mxnet ssd object detection provided in mxnet SSD, while I'm trying with provided data set with two classes, it works fine. But I am trying to use my own data set with 40 different classes, there is a problem.

I noticed that "MultiBoxTarget" always returns zero. I am not familiar with the mechanism of MultiBoxTarget and can not figure the problem exactly.

Is there any one who can describe this function for me? Any helpful hint is appreciated.

K.Dᴀᴠɪs
  • 9,945
  • 11
  • 33
  • 43
Ali
  • 387
  • 4
  • 11
  • Could you provide a sample of your dataset and the code you used? – Thomas Mar 26 '18 at 16:53
  • @Thomas, I tried the "ssd" with the provided trained weights (with 148 iterations) on the link, I could not get the same results as shown in the page!! I cant understand the reason! – Ali Apr 06 '18 at 11:00
  • I have myself not been able to achieve the same map. I got 0.71 after 280 epochs. The reason could be that the data-augmentation is different than the one performed on the caffe example. It is a known issue, you can track it here: https://github.com/apache/incubator-mxnet/issues/9622 – Thomas Apr 06 '18 at 23:15

1 Answers1

1

In regards to reproducing the mAP, the issue Thomas mentioned has been updated and resolved with instructions on how to reproduce the score.

Please see https://gluon-cv.mxnet.io/build/examples_detection/train_ssd_voc.html

from the issue https://github.com/apache/incubator-mxnet/pull/13186.

Vishaal

Vishaal
  • 735
  • 3
  • 13