0
$\begingroup$

I fine tuned MobileNetSSD for object detection using a dataset with just one class (~4000 images). All the training images include at least one bounding box related to that class (no empty images). By following the example with the VOC dataset, the labelmap includes two classes, the background and my custom class. However, as I mentioned, there are no annotations related to the background and I am not sure if there should be any.

Now my fine tuned network performs very well when objects belonging to my class are present, however there are some false detections with very high confidence when the class is not present. Can this be related to the fact that I don't have empty images in my training set?

Improve this question
$\endgroup$
2
  • $\begingroup$ It could be. It would help if you describe your model and its output (how does it do the object detection?) as well as your training data (number of training samples, relative size of object in images).. $\endgroup$
    – Mark.F
    Commented Dec 23, 2018 at 13:48
  • $\begingroup$ The model is the original mobilenet SSD (I linked the paper), where only the number of classes has been modified. I added the number of training images and the link to the mentioned example. As for the relative size of objects, it varies a lot but I don't have precise information about that. I will check and try to provide some more information $\endgroup$
    – firion
    Commented Dec 24, 2018 at 9:52

1 Answer 1

Reset to default
1
$\begingroup$

It is correct that to eliminate the false positives, the training must include images that reflect true negative cases. The objective of training is not this.

Find the objects of class $O$ in scene $X$.

It is this.

Find any objects of class $O$ in scene $X$.

Training is effective when sample $S$ is drawn according to the same distribution as will be used for drawing at test time or during actual use.

It is not necessary to annotate the background if the loss function is chosen so that finding nothing when nothing there is indicated as a zero loss and a proper distribution of those cases exist in the training sample.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Thank you, from a theoretical point of view it is perfectly clear. From a practical point of view, I am not sure whether I should associate background images with empty annotations, as far as MobileNetSSD is concerned. Do you have some information about this? $\endgroup$
    – firion
    Commented Dec 31, 2018 at 10:56

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .