Images with no labels used in FOMO training?

I am trying to train a FOMO model. I have gone through all my images and labelled them. They are a mix of images where the object is present and labeled - and images with no labeled objects in them, but have things that are being incorrectly detected as the object. Are the images without objects being used in training? I remember that some Model training scripts only pull in images with objects.

I am trying to train a Bicyclist detector - but the model sometimes gets confused with cars (also have wheels) or people. Would it help if I also label those objects too? Would it help the model ignore them better?

@lukedc Can’t comment on the accuracy, but all images are used indeed.

1 Like

Awesome! I will keep feeding in examples of “background” images then. It does seem to be improving the model. Thanks!

yes all images are used; we add an implicit “background” class to all FOMO models so any area that you haven’t labelled with your objects of interest contribute towards the models understanding of the background. if you primary problem is false positives then you’re doing the right thing, more instances of the background will help.

1 Like