Localization criteria for object instances and association to ground truth objects

Is your feature request related to a problem? Please describe.
In order to define some instance segmentation related metrics such as panoptic quality, one needs to be able to identify true positive instances and their corresponding ground truth. The first step in that pipeline is to define localisation criteria followed by an association strategy

Describe the solution you'd like
Implementation of the different localization criteria resulting for each element in a list of possible reference instances.

• Boundary IoU
• Mask IoU
• Box IoU (already implemented)
• Distance based criteria of centre of mass (must be lower than a threshold)
• Centre based criteria (if centre of reference is included in prediction)

The hit criteria output is then followed by an association step:

• Greedy based on score - The prediction instances are ranked by prediction score and assigned (if possible according to the hit criterion) to a single ground truth (the one with the best score). The reference associated object is then removed from the list of possible objects
• Hungarian algorithm - Association based on minimisation of overall hit score

Need optional consideration of multiple instances hitting a given reference instances either removing them from the validation or considering them as FP

Start by locating the existing Box IoU implementation and trace how localization results are represented. Define the remaining boundary, mask, distance, and centre criteria, then design the greedy and Hungarian association paths, including duplicate-hit handling. Done means the requested criteria and association strategies are implemented with their optional duplicate-instance behavior specified.