I wanted to know whether not annotating some categories in our images, and feeding them into our neural network, would have any adverse effect on our model in detecting the desired objects.

To better clarify what I mean, here's an example of what I am looking for in my question: Let's say we have a object detector network, which has only 2 categories; "cars" and "persons". We have 10,000 pictures, in which in every image, we have at least one instance of either categories. After annotating all of the "cars" and all of the "persons" in the first 5,000 images, we feed them to our network, and we see we get an AP score of "1.0" for "cars", and an AP score of "80.0" for "persons". ==> Since our model is already doing a great job at detecting "persons" and to save us some time, we decide not to annotate any more "persons, in the remaining 5,000 images, and to only annotate "cars" instances.

==> So, here's again my same question; will this approach have any adverse effect on detecting the "persons" as well? after feeding the next 5,000 images into the network, will this make "persons" AP score to go up or down?

  • $\begingroup$ I assume you're talking about multi-label classification. If some of the images actually contain a car (or something which looks like a car) and the car is not annotated, the model will learn "not a car" and this could affect its performance on cars. The other option is to train two completely independent models (binary classification). $\endgroup$ – Erwan Nov 19 '20 at 14:18

Your Answer

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

Browse other questions tagged or ask your own question.