Using EfficientDet for Training on Unlabelled Images

[khanfarhan10]

Is there any way in which I can use this model to train on a dataset that contains negative images? (images with no objects)??

Is there any workaround?

575

Also how to create such a dataset :

Corresponding Annotation JSON entry :

{'area': 0,
  'bbox': [],
  'category_id': 2,
  'id': 0,
  'image_id': 0,
  'iscrowd': 0,
  'segmentation': [[0]]}

Is this how it should be defined.

[zylo117]

It's possible but you have to modify the code by yourself. And I don't think empty backgrounds really help.

[khanfarhan10]

What I did might just be stupid:

For the empty bg images, I made a fixed BBOX of (0,0,1,1).

Is this right or is probably a very bad idea?

[zylo117]

There is no bg class at all, how should efficientdet classify the anchors with no real target in it.