Labeling > Support providing Annotations as optional Input

0 Kudos

Hi,

I am using the new Labeling recipe as the name suggest to add labels to unlabelled images. In my use case, the images are quite big in terms of dimensions and each images requires around 10-20 annotations.

This means, the manual task of labeling a decent amount of images is quite high.

So far I have labelled around 70 images and used them to train my model. Next, I have applied this model to larger set of unlabelled images with great accuracy but also with more room for improvement.

To easily provide the model with many more training images, (highly accurate) outputs from the previous model runs (label information for detected objects and categories) could be reused as optional input for the labeling recipe working on training images.

As a result, the labeling recipe would support:

  • A folder of images + metadata (list contents recipe) as inputs as it currently does
  • But additionally and optional, a dataset containing annotations for at least some of the images provided by the image folder. Those provided labels should be applied to the labeling annotate process, meaning that bounding boxes along with their categories will be drawn on the images. All the annotator has to do is to either accept those labels (if they are correct) or make minor modifications. But in total, a lot of time could be saved that way.

Labeling_Annoations.png

4 Comments
JIshikawa
Level 3

An alternative and probably better solution would be to join the file listing outputs with scored label data (from a model run) and then feed this "enhanced" file metadata + image folder as input to the labeling recipe.

1_folder_listing_original.png

folder listing: original metadata

2_flow.png

New labeling input flow where the original folder listing metadata is joined with labels from a model run

3_folder_listing_joined_with_label_info.png

updated folder listing containing the original metadata plus label info

   

 

An alternative and probably better solution would be to join the file listing outputs with scored label data (from a model run) and then feed this "enhanced" file metadata + image folder as input to the labeling recipe.

1_folder_listing_original.png

folder listing: original metadata

2_flow.png

New labeling input flow where the original folder listing metadata is joined with labels from a model run

3_folder_listing_joined_with_label_info.png

updated folder listing containing the original metadata plus label info

   

 
MaximeA
Dataiker

Hello @JIshikawa 

Thanks for raising this !

We have logged this idea internally and will let you know if we have any updates

Status changed to: In the Backlog

Hello @JIshikawa 

Thanks for raising this !

We have logged this idea internally and will let you know if we have any updates

MaximeA
Dataiker

Hello @JIshikawa 

We released the ability to input pre-labels to a labeling task in 12.4.0 ! 

These pre-labels should be in a column of your input dataset: 

 

Screenshot 2024-02-26 at 12.08.22.png

 

They will appear in the Annotation view and are editable/adjustable/removable, like any other label. 

Thanks again for the request

Status changed to: Released

Hello @JIshikawa 

We released the ability to input pre-labels to a labeling task in 12.4.0 ! 

These pre-labels should be in a column of your input dataset: 

 

Screenshot 2024-02-26 at 12.08.22.png

 

They will appear in the Annotation view and are editable/adjustable/removable, like any other label. 

Thanks again for the request

JIshikawa
Level 3

Awesome and many thanks!

Awesome and many thanks!