0
$\begingroup$

I'm trying to generate data for my object detection network (which will be used for TensorFlow: ResNet).

What I'm currently curious about is this: if I have the same total amount of data (each data containing coordinates and class) is using an image with multiple objects better or would it be better to have single objects for each image?

For example, would it be better to have 5 images where each image has 10 objects (total of 50 data) or to have 50 images each with a single object (total of 50 data)?

(The numbers are for illustrative purposes only. I would be using a few thousands of data)

$\endgroup$

1 Answer 1

1
$\begingroup$

It's always best to train with a dataset which best matches the test-time distribution. So the answer depends how many objects you expect to see at test-time.

There might also be a small benefit from training on images with many close together objects, even if such scenarios rarely appear at test time, simply because they are more difficult.

$\endgroup$
1
  • $\begingroup$ It makes sense to train on the more difficult, uncommon scenarios, but is there evidence (theoretical or empirical) that it gives better results? I could believe that a model would learn to expect crowded images. $\endgroup$
    – Dave
    Oct 23, 2019 at 1:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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