background images generating masks

[rohanricky]

I am using 2 classes to train : background(class_id=0) & object class(class_id=1). My code is similar to balloons.py code. The object gets detected well but the problem is background also generates masks and it gets predicted as class_id=1. I trained the model for ~50 epochs. All the background images which don't have any object generate masks. How can I solve this issue? @waleedka Can this be resolved by training for more epochs/what have I done wrong?

[fastlater]

And your training output variables (loss, bbox, mask)? You need to check if the model was learning something useful. To run for x number of epochs does not mean model will learn for sure.

[rohanricky]

The model is giving sub-par results for image with object. But it is also giving masks & bboxes for images without objects. I can't train with background images only because they get filtered out: In model.py if not np.any(gt_class_ids > 0): continue

[windson87]

@rohanricky Hi, i have the same issue. I wonder why it doesnt take pure background images to train, coz in many cases, the occurrence rate of instances is not that high to have enough train data. Also in that kind case, isn't it also important to learn images without any instances to avoid false positives, right?

[MathiasKahlen]

@windson87 Did you find any way to train with pure background images? I am having issues with false positives on certain objects in images without any of my desired objects. So it would be nice to be able to train negatives. I'm just not sure how though.

[cooleel]

Any updates on this topics? I saw the same issue, couldn't get rid of these false positive predictions on negative samples.

[reedless]

Been a while but you can check out #1088