2

I have a dataset, where objects are very close to each other. So, the question is: what is the best approach to label them?

There are two possible options:

  1. mark objects so that they will not intersect (it is difficult, surroundings are not included in the label area)
  2. mark a larger area of objects, but labels will intersect

What is more practical?

enter image description here enter image description here

Valery Noname
  • 121
  • 3
  • 2
    I've always seen it not intersecting in examples of CV datasets. it's probably better if you can avoid intersection between labels. Yet I thinks the most important part when labelling is that it does not take you days and weeks to label your data, so you shoudn't worry if sometimes it intersects because you're drawing swiftly around your objects. Having an example of image in your post may help understand the situation and give a more precise answer than this one tho. – Ubikuity May 17 '21 at 12:57
  • 1
    Hello. Welcome to AI SE. Maybe you should clarify what kind of task will your dataset be used for? Object detection? – nbro May 17 '21 at 16:31
  • my dataset is proprietary, I give an idea of the problem, pictures were uploaded – Valery Noname May 19 '21 at 20:44

1 Answers1

1

In my opinion, the second option will be more general. You can refer to some famous datasets for object detection task such as COCO or Pascal VOC, they usually accept the intersect annotations. As the image below, image from this link where they process the annotation of COCO dataset.

I think the reason is that the model will be easier to separate the intersect patterns in the bounding box than interpolate the missing patterns of the object to understand it

elephant

CuCaRot
  • 892
  • 3
  • 15