We have a geospatial project going on LB, where we’ve completed one round of annotations with a specific labeling ontology. On running experiments we found that due to uncertainty in the labeling step, the model performance is not as good. So what we’d like to do is to relabel those same images, but using a new ontology. We’re trying to figure out the best/less lossy way of doing this. The requirements are:
- Dont lose the previous labels, but label again with new ontology
- If possible, view the old labels while re-labeling (as a guide)
Some options we’re considering:
A. Make a copy of the project, then change the ontology - is this even possible / recommended?
B. Add to the existing ontology, but filter out the old labels after export
Your help is much appreciated!
Hello @anika ,
While we do not recommend to modify ontology for existing project (if there is/are label(s)).
You can add/modify, tools/classifications to your existing ontology and do what we called a “delete and requeue with template”.
What does delete and requeue with template do, you may ask.
Basically it send back the asset already labeled with existing annotations (with template) back to the labeling queue.
ref : Data Rows tab
Now having a complete new ontology to an existing project will basically hide all annotations and this is not recommended, this would leads to unexpected result (you may have those annotations in the export for example).
Thanks a lot for this - this is a great way to go, except that i want to try this on a copy of the project, so I leave the original annotations untouched (so they can be used later to compare the 2 strategies). Do you know if there’s a way to copy an entire project, either through the API or the web dashboard?
There is ways to do a project copy via our SDK (you basically need to export your project and import annotations in a new project) but this does not address the main point here, if you modify your ontology you will end up with a broken original project (given you are changing the ontology).
In your case I feel you need to have new ontology and a new project, copy the feature you would like like to keep, import them as MAL to save time and then you will be able to compare without having issue down the line.
We have technical resources to the above steps : https://github.com/Labelbox/labelbox-python/tree/develop/examples#readme
Please visit our Community @ https://community.labelbox.com/