So, I wish to change the ontology for an object that needs to be detected, maybe adding a few other labels or adding sub-classification for an existing label.
We have labeled most of the images till now, but those images did not require these new labels or sub-classification.
This label is required for some new images.
Would I need to rework the old labeled images if I make this change in Ontology?
Or would it be OK since those images had nothing to do with this label?
Hope you are well today.
Please feel free to add any new ontology tools or classifications to your Ontology. I would then re-queue the assets you would like to re-label.
Adding tools to the ontology will not force you to rework all the labels. This action will add the classification or tool option to all labels in the project, but will not be required to fill out or use.
Let me know if you have any questions!
Thanks for your wishes, yes, I’m very well; How are you?
Also, thanks for your response, but I’m still a bit confused so let me give an example for further clarification.
Let’s say I have 3 labels mountain, river, and house
Now I’ve labeled all the images that had mountains or rivers (more than 1000)
Now I wish to upload images that have a house in them, but I wish to add types of houses like small, medium, and large to the labels.
So, this wouldn’t mess up the earlier done mountain and river labels, right?
Also, I wouldn’t have to label or review those previously labeled images, right?
And adding to this question,
In your opinion what would be better;
To have specific labels like a small house, medium house, large house,
Or use sub-classification labels for house (small,large,medium).
Hi @utkarshtomar736, if you add to your ontology, ensuring that you do not change any details about the tools that already exist in your ontology, then you will not have to re-label or review those previously labeled assets.
On the contrary, let’s say you added a required, global classification. In this case, you would then have to go back and answer this required classification for all the assets, because you have changed your ontology in a way that fundamentally altered the requirements of each label. As another example, if you deleted your original
mountain tool, even if you made a new tool also called
mountain, it would have a different
featureSchemaId, and thus the annotations made with the original
mountain tool would no longer be visible. Fortunately, it does not sound like you plan on doing any of these actions, but I thought you might appreciate the additional details as to what changes will have downstream consequences.
As to your final question, this decision depends a bit on what you are trying to train your model to identify. As a general piece of advice, since many training engines do not allow nested classifications and require a “flattened” label, I would suggest proceeding with three different “house” tools (
large_house). When a model is given more detailed labels, you increase the chances of it making a better prediction. Then again, this depends on what you are hoping to identify. If predicting the size of the house is essential to your goal, then you should use multiple tools, but if identifying houses in general is the most important element, it might be wiser to proceed with one
@Zeke_Labelbox Thanks for the detailed explanation and yes, it’s much appreciated.
As for the 2nd part, the house was just for example;
I wish to detect PowerStations, but they are of different types like hydro, thermal, solar, etc.
So, I thought it would be better to have some sort of difference in their label.
But the goal is to identify the PowerStation only not its type.
So would you suggest having,
- One Powerstation label.
- Multiple PowerStation labels.
- One PowerStation label with multiple sub-classification.
I would suggest option #1 – to have one PowerStation tool. If your goal is to identify PowerStations and their type does not matter to you, then you should make your ontology most accurately reflect your goal.
Ok, thanks for the explanation been learning a lot from you.