0

I've started annotating using LabelImg tool and drawing boxes but I have too many pictures in the images (like lots of grapes in the image). For better trained model, is it required to box all pictures or is it okay to leave some?

I am trying to train a Faster RCNN model.

example: enter image description here Thanks

Innat
  • 16,113
  • 6
  • 53
  • 101
RedFox
  • 1,158
  • 2
  • 15
  • 28

1 Answers1

1

I think as long as it's not getting too small bbox and, visually recognizable to the human, or possible to get features within it - that's fine.

For example let's consider the following cases, a dataset contains such meaningless annotation (red marked) which normally an engineer would skip those bounding boxes (box['w'] * box['h']) < some threshold.

enter image description here

Innat
  • 16,113
  • 6
  • 53
  • 101
  • Line in your image, my image has no depth. You see all images clearly – RedFox Feb 15 '21 at 04:52
  • Do you mean in your image, all objects look clearly? – Innat Feb 15 '21 at 05:05
  • In that case, yes, it needs to be annotated as much as possible. – Innat Feb 15 '21 at 05:07
  • Added the picture. thanks. Do you think I can skip partially hidden one? We have lots of pictures to annotate. :-( – RedFox Feb 15 '21 at 05:12
  • Oh, I see. Yes, I think you can skip the partial one. The object looks similar, so the trained model should recognize such an object. – Innat Feb 15 '21 at 05:22
  • follow-up question. If I train a model with 250x250 images, can I use the model to infer bigger images, for example 500x500? – RedFox Feb 15 '21 at 05:26
  • 1
    The test image need to be preprocessed same as evaluation time. So its better to resize image. – Innat Feb 15 '21 at 05:30