1
$\begingroup$

I recently joined a new project, and saw that they are using object detection instead of image classification for one of the business cases. The images can only belong to one class (example, the image is either a cat or something else) and the location of the object does not matter. They just want to detect the presence of it. The training/test data have bounding boxes drawn over the objects, and I am trying to find evidence, if any, as to why this method has any advantages over normal classification. My initial though was that with bounding boxes, provided they are accurately drawn, the model can learn key features better than usual classification, since in classification it is learning the entire image, but I am not sure, because in that case, then most classification problems would have bounding boxes drawn over the key object to improve results.

Are there any advantages of the current approach or is classification the 'right' option?

$\endgroup$
2
  • $\begingroup$ I'm very interested in this topic. Have you had the chance to investigate it further? $\endgroup$
    – firion
    Mar 9, 2023 at 15:08
  • $\begingroup$ @firion unfortunately I did not have time to get around to it, if/when I do I will let you know :) $\endgroup$
    – user27771
    Mar 13, 2023 at 8:01

1 Answer 1

1
$\begingroup$

I think you are right, having bboxes help the model localize objects much better, and usually, the classification comes with each bbox it found so the space now will be reduced to fit just one important object that it is trying to make a prediction, not the entire image where there are a lot of objects.

Classification is only good if your input has information controlled, or else it can cause confusion for the model as the input may not have a single label for the whole image.

$\endgroup$
1
  • $\begingroup$ Thank you for your input. I guess the best option in this particular case would be to implement a classification model and compare the performance of it with the Object detection model to confirm if the OD model does have any benefit. $\endgroup$
    – user27771
    Feb 24, 2023 at 6:28

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .