0
$\begingroup$

I am currently using faster_rcnn to train a set of images with only one category, in fact, there are only good images and images with defect in the whole dataset,and I use roboflow to labeling this datasets,

My problem is that when I encounter a good image, I just mark this image as null, but I don't know if it is effective to put it into the model for training. Should I frame the entire picture into a new category(two categories in total), or remaining mark those image as null?

$\endgroup$
1
  • $\begingroup$ Hi @Kekai and welcome to AI Stack Exchange! It would be helpful to add more details to this question to help identify the details of your problem. Thank you for posting, and we look forward to more of your questions on this site! $\endgroup$
    – DeepQZero
    May 23, 2023 at 14:59

1 Answer 1

0
$\begingroup$

Marking these images as null is the correct thing to do. Object detection models will learn not to predict unlabeled areas (background area) through their loss function.

It can sometimes be useful to add negative annotations if there’s a particularly subtle difference. For example “screw” and “bolt” — this will give the model more information to go off of, let the “objectness” part of the model so it’s thing, and the classification part do its thing also.

By the way, be sure not to add too many null examples to your dataset or it might learn the optimal strategy to optimize its score is “predict nothing”. If this is happening you might want to use the “filter null” preprocessing step in Roboflow to adjust the ratio.

$\endgroup$
0

You must log in to answer this question.

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