Images with no object

Using Labelbox
#1

Hey folks,

I’m new to using LabelBox, and the task I’m using it for requires me to split satellite imagery into tiles and annotate the required object.

Now the issue is there are quite a few tiles that don’t have any object to annotate, I want to know what can be done with those images.
I do want to keep the annotated data for those images like those would not have any bounding box;
But I am unable to submit those images without annotating the image.

So, is there something I could do like can I create a classification radio button of no-object and click yes for such images?
Or any other workaround.

Any help or feedback would be much appreciated

#2

Hi @utkarshtomar736, I can think of three options:

  1. Use tiled imagery format. Using this format, you can skip converting geospatial tiled imagery into image chips. Geospatial / tiled imagery. This option is available in paid subscription tiers only.

  2. Skip images that don’t have any objects to label.

  3. Use similarity functions (powered by embeddings) to create a batch of images that contain objects of interest. Using similarity function, you can find similar images that contain or do not contain certain visually similar features. Learn more: Similarity functions

1 Like
#3

Hey @manu,
Thanks for your reply, I have a few concerns

  1. I do have access to paid tier and geospatial/tiled imagery (I think), would you say this would be the preferred way if I did have access.
    Actually, this looked a bit complicated as I did have the images already tiled to Png format I didn’t bother to look much into it
    Like it uses its own map and stuff so I was confused if it’s for data collection as well,
    Elaboration on this technique would be greatly appreciated.

  2. I had a doubt with ‘skip’ as well, like what does it do;
    Does it mark the image as one with no object or does it push the image to be labeled later?

  3. I would definitely look into the Similarity Function.

Thanks again for the guidance, and clarification for the first 2 points would be greatly appreciated.

Edit: For context, I used Landsat-8 imagery (google satellite) for data collection and got images with various zoom, etc.
Would I have the freedom to do so with Slippy maps?
Sorry, quite confused about how this method works.

#4
  1. Yes, that’s the preferred way to label geospatial data. You should be able to import landsat data by specifying the skippy map URL. You can also specify alternative tile layers such as from open street map or google maps. Example: Geospatial / tiled imagery
  2. Skip is essentially a global classification on the data row. Once a labeler skips a data row, it is marked as such and removed from the queue. You can easily access skipped data rows any time.
1 Like
#5

Thanks again @manu,

I would definitely look into geospatial imagery; I think I understood it the wrong way.
Also, thanks for the clarification on the skip.

Hey, the forum and your replies have been very helpful and I’m sorry if I sound insensitive but I might require some clarification on an urgent basis when I go ahead with these approaches;
Would there be any Moderator, Admin, or Support team for Labelbox with whom I can get in touch in case we need some help?

Just to be clear, I truly appreciate you taking your time out and replying.

#6

Anyone from Labelbox can reply here. If you want to share sensitive information about your data or account, reach out to support team. Learn more: Customer support

1 Like
#7

Thanks again, you’ve been a great help.