Label assistant also display polygonal boxes


I’m in the process of converting my dataset with polygons boxes.

Is there a way to support annotation with Roboflow model?

So far the lable assistant only delivers rectangular boxes.

Or do I have to adjust something during training so that the polygon boxes are also displayed?

Yes, Roboflow does support polygon annotations. Polygon annotations provide more specific information about the data you’re using to train a model, which can lead to increased training performance and results, and also improve data augmentation [source].

To annotate images with polygons in Roboflow, you can use the Polygon Tool by clicking to add an initial starting point, then continue clicking to add more stationary points around the object, and double clicking or clicking the original vertex to close the polygon. The shortcut key for the Polygon Tool is “p” [source].

Roboflow also offers a feature called Smart Polygon, which can generate annotations by clicking an object of interest. There are two versions of Smart Polygon: Standard, which is ideal for small items, and Enhanced, which is ideal for most use cases [source].

However, it’s important to note that annotating with polygons takes more time and bounding boxes are usually sufficient to build an initial prototype or model for assisted labeling. Once you have an initial model working, you can go back and improve performance later with polygons, by relabeling your images, if necessary [source].

Hi! I’m Roboflow’s AI powered bot. This is an automated response powered using AI with knowledge from our docs, blog, and forum. If this solves your issue, please mark it as the solution using the checkbox icon. If this doesn’t solve your issue, please follow up with what you’ve tried/why, and the community will continue to chime in as usual.

Does this mean that if I have labeled all images with polygonal boxes and then do a training, the label assistant will label new images with polygonal boxes?

I have now done a test. I created a new project and labeled the images with polygonal boxes.
Then I created a Roboflow model and uploaded new images.
However, the label assistant only labels with rectangular boxes and not with polygonal boxes.
Why is that?

Hi @staebchen0

I want to clarify some things to get us on the same page so I can help you better. You are looking to train an instance segmentation model (which produces polygon predictions) on your dataset, right?

If you have an object detection model, (which produces bounding box predictions) there isn’t a way to force it to produce polygon annotations, since they are different model types.

In an object detection project, training a version will result in you getting an object detection model, regardless of the polygon annotations you make (they will be converted into bounding boxes).

If you’d like to train an instance segmentation model, you should create an instance segmentation project.

Hi stellasphere,

Thanks for the info.
At the moment I have an object model and use polygonal boxes for the label marking.
Could I put my dataset, which is already labeled, Converter into a segmentation Model?

I have created a new project and selected Segementation, unfortunately the labeled images from the other project cannot be transferred. Although I only used Polygonal label

it would be very tedious to relabel all the pictures again

Hi @staebchen0

Polygonal boxes usually refer to a 3D object, so I want to clarify, by polygonal boxes, you do mean polygon shapes, correct? If that is the case, you can indeed create a new segmentation project and transfer your images into that project.

I can see your attempt was through Roboflow Universe’s clone feature, which doesn’t work, but you can do this by creating a new version of your datasets (make sure to disable and remove any/all preprocessing and augmentation) and exporting it, then importing it into your new project.

Hi stellasphere,

yes, I meant polygon shapes.

Which export format should I choose? Or does that not matter for the import?

and how should I then upload the dataset with the labels? Do you have an example?

I think I found a way.

Thank you for your support :blush:

Hi @leo
I have another question

why are the images all saved with image_jpg when uploading? In the Coco dataset they are stored with Image.jpg

Hi @leo ,

There still seems to be a problem with the import. There are images with the same name in the original dataset.

This is now causing problems when importing into the new project. The labels are taken from a duplicate. Of course that doesn’t fit.

How come?

Hi @staebchen0

For the export format, as long as it’s a format that we support imports to, it doesn’t matter. I recommend COCO, just because it’s a common format that we support imports from.

That works too, but you can also un-zip the dataset file and drag-and-drop it into the Upload page!

I’m not sure about that. Is it causing a problem?

Could you share your previous project’s Universe link? (or workspace & project ID if it’s private)

I have now deleted all the images with the same file name. As I said, it should be prevented when uploading that files with the same name can be uploaded. This can happen quickly with large data sets

This topic was automatically closed 7 days after the last reply. New replies are no longer allowed.