1
$\begingroup$

In the paper SSD: Single Shot MultiBox Detector by Liu et al., 2015, the Matching strategy section reads:

During training we need to determine which default boxes correspond to a ground truth detection and train the network accordingly.

Now, suppose we have some cases in our data which do not have any object in the image (hence, no ground truth bounding boxes). How do we handle this situation during training?

Improve this question
$\endgroup$
2
  • $\begingroup$ Machine learning models which can learn from both labeled and un-labeled data is semi-supervised learning. $\endgroup$
    – Sycorax ♦
    Commented Jun 18, 2018 at 13:18
  • 1
    $\begingroup$ @Sycorax As I understand the question, it does not regard unlabeled data, rather data which are correctly labeled as having no objects in them. I removed the [semi-supervised] tag and returned conv-net again, as both referenced network architectures are conv-nets. $\endgroup$
    – Jan Kukacka
    Commented Jun 19, 2018 at 7:33

1 Answer 1

Reset to default
5
$\begingroup$

You need to remove those images from the training set.

Since these networks work in supervised fashion, presenting them with images without any bounding boxes would either have no effect (just wasting the computational time), or even be inferior to the performance (since you would be learning the network to NOT predict anything even if there were real objects in the image).

Improve this answer
$\endgroup$
2
  • 1
    $\begingroup$ At test time, I may get images without any objects in them. Don't you think I should train the model to generalize well for such cases too? $\endgroup$
    – Rishabh Agrahari
    Commented Jun 19, 2018 at 6:13
  • 1
    $\begingroup$ I assume in the rest of the training data there are images which have objects in part of the area and nothing in the rest, right? Then the network is learning to predict objects where they are, and predict nothing everywhere else. It only takes the content into account - there is no need to explicitly tell the network that images without objects may also happen. Unless you struggle with false positives, I would leave these out. Note that some open implementations of SSD will not work if there are no bounding boxes in a training image. $\endgroup$
    – Jan Kukacka
    Commented Jun 19, 2018 at 6:27

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.