When I use an image in which there are targets to train a detection network, I can use the xml annotation to get the idx and rec file. But when I want to use an image in which there isnot any targets, I do not know how to do. For the xml, it need the position（xmin,xmax,ymin and ymax）of a target. So there are no x or y for the negetive image to generate the xml annotation.
And I also tried to write an xml annotation like others except the object part, it is not work.
How can I use the negative image to train the detection network like SSD?
@zhreshold
I guess you are talking about the xml file defined at https://gluon-cv.mxnet.io/build/examples_datasets/detection_custom.html Can you try putting 0 in all xmin, xmax, ymin and ymax for negative images? Does it work?
Yes, the xml file is defined like the url you gave.
And I tried the way that setting 0 on all xmin, xmax, ymin and ymax. The network can run but in fact these negative images do nothing to the results. From the detection results, I see there is not any performance improvement.
So, I think this way is not OK.
Do you have some idea else for the problem?
Thanks for your reply and I am very looking forward to it.
Unless MXNet team fixed this, I would suggest “hacking” the issue by generating the synthetic dataset with the use of both “positive” and “negative” samples in the following way. Keep the annotated images as they are, crop some boxes out of them and embed crops into the negative images, say, one crop per negative image to some random position. Everything outside the crop should be considered as the background, that you probably want to achieve.