2

I'm trying to find free multi-instance single-label datasets for object detection online.

By "multi-instance and single-label" I mean that each image contains only object belonging to one class, but can contain more than one object of a certain class.

I found a lot of datasets for multi-label, but none for single-label

Any ideas are highly appreciated,

thanks in advance

Jürgen K.
  • 71
  • 3

1 Answers1

0

Do you mean that several objects in an image aim to one specific label?

For example, if you have an image with an astronaut, buildings, a car, and a rocket, it has the label "Space"; and an image with buildings, cars, and a circuit has the label "Racing"?

In that case, you would have to add one step to multi-label by making sense of several labels into one thanks to a logistic regression or something similar.

In addition to that, here is a publication that might help: https://arxiv.org/abs/1811.08400

Nicolas

Hedayatullah Sarwary
  • 414
  • 1
  • 4
  • 11
Nicolas Martin
  • 4,509
  • 1
  • 6
  • 15
  • This would not be object detection type of a scheme. What you mean is done for image classification only. What I need, to adapt your example, is images with just astronauts or just buildings or just cars with their bounding box information. – Jürgen K. Jun 21 '21 at 07:46
  • Sorry, I would need some information to aswer correctly. Can we consider your problem as a subcategory of multi-label? Normally, if you use a multi label solution, you should be able to get single labels as well. Do you want to detect always the same object in different pictures, or do you want to detect different "single objects" that are unique in different pictures? In this example, you can see that single objects can be detected in a multi label way: https://softwaremill.com/counting-objects-with-faster-rcnn/ – Nicolas Martin Jun 21 '21 at 08:09
  • Single-label multi-instance means that there are multiple object belonging to the same class in each image. Extracting single labels from a multi label solution does not fit well, because the objects are not disjoint. For example you can extract annotation of a person (regarding the image in the upper right corner of your reference) but there still would be makant objects without bounding boxes in the same image. So yes, I want to detect same (multiple) objects in different pictures. – Jürgen K. Jun 21 '21 at 10:50
  • In that case, why don't you use a second label of objects to detect wrong objects that could be mistaken for the object you are trying to detect? ( = "non object") It should improve the learning, you can also correct wrong positives, and you would just have to ignore the second labeled results. – Nicolas Martin Jun 21 '21 at 11:57
  • This is probably a workaround, which still isn't perfekt, because on one hand in mixed-label images the quantity of a specific type of object is usually really low. On the other the "non object" is quiet synthetic, since sometimes it is hard to decide if the object is non type or background already and depends only on the labels that the dataset publishers were interested in. A dataset where only object belonging to one type would be much more interesting, that is why I search for such dataset. Your idea with counting topic was great. Probably there are some related datasets with single-label. – Jürgen K. Jun 21 '21 at 12:15
  • 1
    You can also try an autoencoder, and consider your problem as an anomaly detection, as explained in this conversation: https://datascience.stackexchange.com/questions/55924/cnn-for-checking-existance-of-single-label There could be interesting solutions using masks: https://towardsdatascience.com/single-stage-instance-segmentation-a-review-1eeb66e0cc49 About data, you have the ship detection challenge (=single label) that might be interesting: https://www.kaggle.com/c/airbus-ship-detection + plenty of shared code. – Nicolas Martin Jun 21 '21 at 13:45