I have a few questions about "instance" labels in output file(.json)

Hi I have a few questions about "instance" labels in output file(.json)

1. Is the "instance" label included in the evaluation factor?
2. Is the "instance" label the same as the "class label" used for training?
3. How should I map the "instance" label to the output file(.json)? Does a mapping table between "instance" and "FDI system" exist?

Thanks

Hi,

Instances are used to indicates vertices that belong to the same tooth (object). We try to do instance-wise teeth segmentation. In other words, instances cluster vertices in separte teeth.

1- In the evaluation, we use instances to be able to separte teeth instances. However, the value of the instance is not important as long as vertices belonging to a same tooth are attributed a same instance value. For example, if in the ground truth for a given tooth vertices the instance is equal to 3 but in your result the same tooth vertices are attributed and instance equal to 4, it is not an problem.

2 and 3 - No, you don't need to infer the instance value in your model like for the class label. You need instead to attribute a unique instance for each separate tooth. from a range between 1 to number of detected teeth. 0 is reserved for gingive (see Guideline in github repo).

Best regards Achraf