False Positives

[mancubus77]

I trained a model, and it has a good AP50, but many false positives. I tried to include unlabeled images to the dataset (cfg.DATALOADER.FILTER_EMPTY_ANNOTATIONS = False) , but it doesn't' help much and I'm keep getting same false positives.

Just wondering how you guys overcoming false positives in Retina and RCNN models?

[mancubus77]

COCO dataset looks this:

    {
      "file_name": "images/IMG_6740.jpg",
      "height": 3024,
      "id": 1000135,
      "width": 4032
    },

With no annotation linked to the image