Multi-object in image keypoint labeling and export

Label Studio Support
,
1

I have images for training, and let’s say that each image I am using for training contains multiple boxes. I want to label box corners (8 in total) per box, and like said I’ve multiple boxes and on output I want to have keypoint label to each corner consistent, like top-front-left, top-front-right, top-back-left etc.

example image from net

image
image607×390 34 KB

I do not understand how I could configure label-studio project/task labeling-yaml configuration. Naturally I can create keypoints for top-front-left (tfl), but how do I get box1/tfl and for another box box2/tfl and another.

Or do I need to label single image somehow multiple times, each “round” with just one box. How that is done? Re-uploading same image multiple times?

What version of Label Studio you’re using (for example, 1.10.0).

“latest” available at Docker hub

How you installed Label Studio (for example, pip, brew, Docker, etc.).

Docker

Your labeling configuration.

I dunno - this is what I am asking how to keypoint label multiple object in the single picture. I can add bounding box or segmentation mask per box, if via that and “relation” I can then add mandatory named box-corners (1-and-only-1 per corner per box).

I understand this is exported as JSON and JSON examples seem to have link to object. But how do I configure the labeling project for this is the question.

2

Hi, it sounds like you need groups for your keypoints. You can drag and drop regions on the right panel to create a nested group.

3

Thank you for the suggestion. So there is no configurable object based keypoint, like this keypoint is for box-object and this for cylinder-object etc.

I mean in Theos AI it looks you can create classes Person and Car and have class specific keypoints. So I was thinking that LabelStudio could support something similar.

I see some mentions about multitasking in Yolo, so I was wondering if Labelstudio could be used to label/annotate segmentation mask and keypoints on same training set. So I was hoping that keypoints could be somehow linked to mask/box.