0
$\begingroup$

Computer vision is highly benefited by AI algorithms. Image data is abundantly available. There are different varieties of tasks such as image classification, prediction, segmentation, generation, etc.

Although the collection of the folder(s) of image(s) is mandatory, it may not be enough. Different types of annotations are used in datasets. Annotations can be treated as some extra information related to each image that helps for the AI algorithm under consideration. I want to know the kinds of annotations at the individual image level that are generally used. Although the necessity of a particular type of annotations depends on the task under consideration. I want to know the requirements for the contemporary prevalent tasks including classification, prediction, segmentation, and generation. You are encouraged to provide for more tasks if you are aware.

I know the following types of annotations:

  1. Bounding box(es)
  2. Label

What can be the other kinds of annotations used for images in image datasets?

$\endgroup$
6
  • 1
    $\begingroup$ @nbro can you please check whether this question is on-topic? $\endgroup$
    – hanugm
    Dec 30, 2021 at 9:41
  • $\begingroup$ I think it's on-topic because you're asking about how data is usually labeled (so prepared) in order to train ML models, but maybe a bit too broad. This type of question makes me think that you're curious about different tasks in computer vision and how data needs to be labelled for each of them, but what if there are millions of slightly different tasks? Of course, I don't think this is the case, but I see this potential problem. So, if you could somehow narrow the scope of this question, it would be better. $\endgroup$
    – nbro
    Dec 30, 2021 at 9:58
  • $\begingroup$ Maybe when you write "contemporary prevalent tasks." you could specify which ones. I think these would be "image classification, image detection, image segmentation, and semantic segmentation" (but what if there are more?) and I guess you're interested in supervised learning, so it may be worthing clarifying that too. Again, this type of question is a bit open-ended, which is an aspect that I dislike, as it can lead to many different answers. $\endgroup$
    – nbro
    Dec 30, 2021 at 9:58
  • $\begingroup$ True @nbro I also feel it is a little bit open since the tasks can be many. And many answers are possible. But, I think that these sorts of questions can be easily answered by experts as they see many available datasets and annotations available for them. $\endgroup$
    – hanugm
    Dec 30, 2021 at 23:30
  • 1
    $\begingroup$ If many answers are possible, it's too broad, and can be closed as such. So, please, narrow the scope of this post or at least be more precise about what kind of answer you're looking for. You cannot expect us to list 100 possible labels, if there are 100 different tasks. What do you want to achieve with this question? $\endgroup$
    – nbro
    Dec 31, 2021 at 9:02

0

You must log in to answer this question.

Browse other questions tagged .